Learn about NexusCon!
Happy Thursday!
Welcome to this week’s deep dive exclusively for Nexus Pro members. It’s an honor to have you here. This deep dive is a follow up to my recent conversation with Jean-Simon Venne, CTO of BrainBox AI. I learned a lot from this conversation and want to share my takeaways and the full transcript with you below.
In case you missed it in your inbox, you can find the audio or video here:
Nexus site | Apple Podcasts | Spotify | YouTube | Add to other podcast apps
Enjoy!
—James
Disclaimer: James is a researcher at the National Renewable Energy Laboratory (NREL). All opinions expressed via Nexus emails, podcasts, or the website belong solely to James. No resources from NREL are used to support Nexus. NREL does not endorse or support any aspect of Nexus.
Well, first, let’s just agree that BrainBox AI is a slick startup.
They have the AI expertise, great marketing, a solid business model, new venture funding, a solid sales force, and leaders like Jean-Simon. I especially love the sales philosophy: they’re like a cable provider without a long term contract. You’re not getting married and can cut the cable at any time.
They’re also in on the all-star game approach… AI is their position and special sauce. They’re fine with that living on someone else’s hardware and being white-labeled and don’t need to sell the whole stack. Watch out for this crew!
My main negative reaction: I’m skeptical that this solution can be implemented in a silo, and that seems to be the message I’m getting, not from Jean-Simon, but from others at Brainbox AI. If the building loses connection with the cloud or if the data is erroneous or if the valve is physically stuck, the AI stops working on that section of the portfolio. While it might be tempting to try and package this up as a stand-alone solution, it’s only as effective as the rest of the system. So I’m worried that their sales messages could mislead building owners into thinking they just turn on AI and the building is now as efficient as it can be.
Another thought: there seems to be quite differing opinions on whether cloud-based supervisory control is feasible for buildings. Terry Herr, our next podcast guest, says it won’t work. Same with Troy Harvey. The GSA demonstrations of BuildingIQ had huge issues with it. Jean-Simon doesn’t seem to be worried about it. Soon I’ll be partnering with another engineer to share a new paradigm for supervisory controls—he’s not too worried about it either.
Finally, am I the only one that’s curious about the amount of energy this sort of AI requires in the cloud? It doesn’t do any good to save energy in the building by eating up an equivalent amount of computing power in the cloud. How do those two numbers compare?
What do you think?
[00:10:29]
We realized that if we're able to predict the future of what will be happening in that building with a super high accuracy or about 99%, we basically know the future. And when we realized that the deep learning algo were, were giving us that accuracy for hours and hours ahead, so we knew that in three hours, that room will be too hot. It will hit the cooling set point in three hours, in four hours, in six hours, and we realized that we had that, that luxury to basically know the future of that building with a very, very high accuracy. You went like, wow. I mean, we could do more than continuous commissioning.
We could basically design our control strategy right now to have a better future. And it's, it's kind of going into this, this concept of, when you think about it, all of our HVAC right now, they're reactive. And even if you use a PID loop, you're reactive. So there's something which is happening now, and you're reacting to it. But you're already into that climb or in that descent, you're already going through that event. You're not into the past or looking at the future. You're really into the present. And all of our HVAC systems are reactive. I mean, they are programed which if this value trigger this action until you have this other value, then you could release or you could stop whatever action you were doing.
The entire logic of our programmation is like that, is reactive. So it's kind of similar, then, if you were driving a car and I was blocking your windshield with like a piece of wood or whatever. You're basically allowed to look on both side. So you see what's happening right now. You're allowed to look in the past. You have your rear mirror. But you don't, you're not allowed to look in the front of the car toward what's coming at you. So you'll be going through all kinds of surprise, and you're going to react to the surprise, but you're already going through that surprise. You don't anticipate.
So if you were driving like that, it would definitely cost you more energy because you were not able to anticipate, should accelerate or slow down or just let it go. And it will be very, very uncomfortable for whoever passenger is with you in that car. And I will not mention the damage that you're probably going to create on your car.
When you think about it, all of our HVAC systems are acting like that. Because we don't know the future. So we're, we're following the schedule. We're following actions that are happening already, and we're reacting to these actions by a reaction. And we're trying to stabilize as we go. So knowing the future, suddenly we could start to do preemptive action that would either improve the comfort, or reduce the quantity of kilowatt that our spent to react to an event.
[00:16:47]
Doing the control is the hard part. Because the big issue we have, like with building, when you think about it, is you're not trying to resolve one linear problem. So let's say one linear problem could be, I want to maintain the temperature all the ways within my dead band of the set point. So that would be like a linear problem and you would not need neural network to do that. A classical machine learning tool would do it. But then you want to do, okay, I want to do that, but at the same time, I would like to do it with the minimum quantity of kilowatt hour that I'm spending over time. Oh, but then you have a second problem that you're trying to resolve and you're trying to optimize both of these problems at the same time.
But then, Oh, wait a sec. This is not it. Because the utility, they're charging us on the power factor too. So even though I am saving kilowatt hour, if I'm creating a super peak on the power side, I'm not going to win at the end of the month when I'm going to receive that bill. Actually, that could create a worst case, because the power factor of the bill, it could be very intensive in dollars. So we want also to optimize that kilowatt in power, as a third line that you want to optimize at the same time. So, Oh, this is great. So then we're starting to have a more complex problem that you want to resolve. So you want to optimize these three lines in parallel.
But wait a sec. We don't want to create some cycling on the equipment. Right? Because we will be slowly destroying too fast like a pump or a ventilator. So, so I want to make sure I'm not cycling anything here. I'm treating his equipment with respect and I'm not creating equipment problem down the road. So, Oh, that's a fourth line that I need to optimize in parallel to the first three.
And then your problem is becoming so complex. It would be resolvable. I mean, we would put like probably, you know, you, me and a couple of other engineer, a couple of stats model around a table. We're going to crunch this, maybe with MATLAB, and we would come with the solution. Probably going to take us a few hours. But we would come with that solution of what is the optimal control for the next hour. By the time we do it, of course, that hour is long gone. So whatever we came up as a solution is too late, because we should have done that in half a second if you want to be efficient. And just imagine the cost of all of us around that table. And we would have, of course, to work 24/7, day and night because you know, we need to operate that building and that's only one building.
So it's, it's all feasible by hand, but the power of like a deep, reinforcement learning. And that's the same, the same thinking than the video game. So when the AI is playing against us on a video game, it is quickly understanding what are the rewards, so how to win the game. So if I want to win the game, I need to maximize basically my points and to get points, there are several ways you could get points, right? So quickly the AI is understanding how to get these reward, and how to score maximum value and win the game.
So it's the same, same technique we're using, but we're kind of presenting to the AI as a game, saying we have all of these paths and you need to optimize all of them in parallel. And if you manage to do like a global balance, you're going to win the game. And that's what we call the deep reinforcement learning. And this is what we're applying. And it's producing a better control strategy than the typical control sequence, which was already there in the building. Which is, as you described earlier, you know, based on fixed schedule, which are not always tweaked the right way, which is based on reaction. So it's reacting. Remember, the AI knows what's coming, so he's got a clear advantage on that control sequence because he knows what's going to be the demand for let's say for the chiller in two to three hours. So it could pre-produce water without creating a super peak.
[00:26:03]
So one of the big challenges we had to resolve-, and it was nothing to do with artificial intelligence, was the fact that there is a lot of different controllers out there on the HVAC side. We actually did a count and we went all the way up to about 700 HVAC controlled protocol in the world. Some of them are, you know, were created by company that don't exist anymore. They're not being supported anymore. Some of them were bought by a bigger company and they're trying to support the line and at the same time convince a customer to, to switch over. There's also a lot of different versions, so, so on each of these control protocol there is different version, and a building may necessarily upgrade their control systems. They try to, you know, go as long as they can without upgrading because they want to avoid the expense.
So it's a pure nightmare. You want to interface with that kind of ecosystem, and some of these protocols are completely closed. I mean, there's absolutely no way you could understand that language. There's no documentation. The company does not exist anymore. Good luck to find somebody that knows about it. So we had to create an edge device, a physical edge device. That we could install in the building, and would be able to connect to that existing control system and talk the same language.
So of course, you know, BACnet was the easy one. But think about LonWorks. Think about Modbus. Honeywell is an interesting protocol just to name one. So it's, you need to be able to connect that box and that box is able to talk that language to first of all, discover what are the points, so an auto discovery. Then read the point. And when we read the point for the first time, let me just tell you that it's a, it's a very rare exception when we hit the building, which is already Haystack tag. Most of the building, it's a very weird nomenclature, which is sometime created by the technician that did that set up. We see building, which is, let's say, half in English, half in Spanish. And we see building where people are giving names to the equipment. I mean so instead of calling a fan a fan, they give it the name of a person. You know, this is Andy and it's a fan, but they don't say it's a fan, just this is Andy. So we have to figure out that Andy is a fan. So, you know, nomenclature is a very creative world in the building HVAC control. And so that's a step that, that we need to do manually.
And actually we're working with the NREL in Colorado, and we are creating what we call the Autobot. So the Autobot is a piece of AI which is doing that mapping, that conversion of whatever name they were using in that building to a standard, Haystack tag nomenclature, because it's only when you organize your data in that fashion that the artificial intelligence could understand what it is.
And that is the step that we call mapping, so still manual process in the motion to becoming AI too in terms of the mapping. So if the AI could do about 80% of that mapping, and we still have human doing 20% of the mapping, I think that would be very, very happy if we reached that level of applying the AI for the mapping, naming conversion.
So once that step is done, this edge device connected to the existing controller of the building will start to read the data and send it to the cloud. So we have different way to do that, in term of secure connection. And that reading goes in the cloud where it accumulate in a database, which is specific for that building.
And then we have to wait. We basically have to wait that period of time to accumulate enough data. We're also in the cloud mixing that data coming from the building with detailed, detailed whether data that I mentioned previously. So you want that, that special weather data, which is giving you the wind, the wind direction, the wind gust, the cloud thickness. So we're not just talking about humidity and temperature here. We're really talking about detailed weather, because there's a lot of correlation between what's happening inside the building with sometime driver, like the wind direction and/or cloud thickness, because that's giving you directly the solar radiation intensity, that cloud thickness.
So that weather data is accumulated in parallel on the same timeline than the data points were taken for the building. During that period of time that I mentioned, you know, five, six, seven, eight weeks, in the same season. And it's only then that we have enough data that you could start to apply this prediction.
And for us, this prediction step is very important because it's giving us the quality control that, yes, we could start to do automatic control because the prediction is good. And it's also something you want to keep doing in terms of prediction analysis. Because when you see a degration of the prediction, it's your signal that you should also retrain your AI.
So that's happening in different situations. It could be happening when there's a season change. So you were, you were training yourself, you were operating during the wintertime, and then you were getting in the spring, you're going to start to see new behavior happening in terms of the weather, of course. And that will have a different impact. Switching from heating to cooling is an interesting aspect, and it's happening in the reverse order in the fall, that requires a retraining of your neural network, so they discover new behavior that's happening all along during the first year.
But it also could be a user behavior change. So a tenant is leaving the 10th floor. He's going to another the building. So suddenly the 10th floor is empty, and then then they will do construction on the 10th floor for the new tenant that's moving in. And then there's going to be a new tenant. Right now we are seeing in the COVID-19 crisis, all kinds of behavior change on a tenant side. So there's a tower, which on 30 floors, there's only two floor, which are still occupied, and it's occupied by a government department, which used to work nine to five, now they're in crisis management, so they're working from six in the morning to midnight. So and all of the other floors are now empty. So who's going there to change the entire control sequence to adapt to this new reality of that tower, right? Feasible. Is it being done? Well, the AI automatically recognized there's a big shift happening and that night will retrain itself to this new behavior that it's understanding. And after a few days, we'll be now completely understanding the new setup of the tower and we'll be trying to optimize that new set up.
Well, that retraining is happening in the cloud. So most of the AI is happening in the cloud. And then we have a lot of what we call control algorithm. So we have about 20 something of these algorithm , and they all focus on one part of the system. Either they focus on the entire tower, like we have an algorithm which is optimizing the schedule. So what time should we start the system this morning? What time can we stop the system in the evening if there's a schedule, if it's that type of building. We have other algorithms which are trying to optimize the temperature of the water, of the hot water being supplied into the system at any given time. So same thing on the chiller side. We have other algorithm which are focusing on power peaks. So how can we manage to not create another peak of power that month in the tower.
We have algorithm which are focusing only on air handling unit. So when is the optimal time to start the cooling stage one or cooling stage two? We have other algorithm working only on fan speed if you have variable speed drive. So what is right now my optimal static pressure in that duct at any given time.
And we have to make all of these algorithms work together. So, we need a coach, which is making sure that the team is doing a good teamwork and not working one against the other. So that's the complexity of it. So most of these tactical algorithms are located at the edge device, because the ones that are doing these decision on real time are located at the edge device, but they work in tandem with the crunching happening on the cloud side.
James: So another type of challenge that I thought of as I think about my average building operator, so there's two challenges and they're both pride related. So, when I think about a building operator, I think about one that either has pride in his control sequences or one that has pride in the way that he operates the building without his control sequences, right? So, in the former, it's like, we have, the state of the art control sequences. Our mechanical engineer did a great job. I don't really understand them. I just know they're working really well. And the latter, it's I don't trust control sequences. I operate my building. Right? And how do you guys approach those two separate pride challenges?
Jean Simone: [00:48:37] Yeah, I mean, you're touching probably the most difficult part of any project, right. How do you make the AI work with humans? So, I don't know if you had the pleasure to sit in an autonomous car level four, but, I was told that, I was really not a good person to do that test, because I'm fighting with the AI. I want to control the wheels. I don't trust it, and it's because I've been conditioned with, you know, a lot of years of driving myself. And for me, you put me behind a wheel and I will drive, right? I'm not gonna just cross my arm and look at the show, right?
So, I think anybody, which has, you know, been maintaining a building for, for years and years, and it's becoming your building and you're very proud of what you do, and you should be, it's not an easy job. You know, I think of all these tenants which keeps complaining and you know, issues, and you're trying to maintain that level of happiness, with all of the users of the building, it's an art. So with the tool that you were given, you pushed them and you configured them through a level of which you're quite satisfied. And you did put so many hours to make it, tweak it so it does the best it could with whatever you have. So, and then suddenly you have that piece of AI coming in and saying, Oh no, we're going to go get like a lot more value out of this building. And it could be very frustrating, to see that, wait a sec. You know, do you know how many hours and years that I put into this?
So, and there's an interesting story about this. I was talking to one of these person that we're describing here, which spent his entire career in one tower. And he's about five years away from retirement, and he's the ultimate expert of that building. And he was looking at me, he said, there was nothing the AI could do better than me. And I was like, okay, yeah, interesting discussion. And I said, well, are you going home at night? And he said, yeah, yeah. Yep. Seven to three, sometimes 3:30. Okay. Perfect. So who's-, if you're not here, what happens? Because there's a shopping mall at that bottom of that tower, during the evening, you know, Thursday, Friday night, that's open until nine. You know, what happens in the shopping mall?
And he said, well, if there's a problem, they're going to call me. Okay. And you like that, to be called? No, no. No. Okay. And what happens Saturday? Shopping mall open, right? Yeah. Yep. They will call me if there's a problem. And what happens if there's something happening during the night and then? Oh, I mean, that's not good. So what about the AI is becoming your second in command, and it's managing the ship 24/7, so you could have a better night to sleep, so you, you could go on vacation two weeks somewhere in Europe. You could go for lunch, not being stressed that your pager is going to ring. And he looked at me and said, okay, now I understand.
And he looked at me and said, and maybe if I show the AI a lot of things in the next five years, maybe I could retire with peace of mind. And I said, yeah.
James Dice: [00:51:26] Yeah. And the building owner would have peace of mind as well.
Jean Simone: [00:51:29] Yeah, so, the AI is not here to replace people. AI Is here to give us the ability to do more on our daily basis. It's increasing our capability or capacity to do a lot more and to maybe focus our creativity on the most important things and letting all of the basic stuff being done by the AI, freeing you up some time to do more important things with your expertise.
[00:56:56]
It's interesting because we were talking about that a few weeks back, trying to understand what will be this new world. Last year we were selling more on we will improve the comfort, so, you know, we will reduce issues you have on the comfort side. And a lot of customers were like, it was their key priority. Yes, I want to save money, but I have this tenant on the sixth floor, which keep complaining and we can't stabilize the temperature. And if your AI could fix that, it's worth the money. And there was also a lot of, let's save the planet. If we reduce the energy intensity, I'm going to be able to say that my building consumed less per square. I'm helping the planet. That was really the spin on the sales side.
What we're seeing now is that kind of completely shift now. Now people are, okay, I need to save money. I lost two tenants. They're not paying their bill anymore. My revenue are coming down, so now I need to do whatever I can on the expense line to reduce that one, because the bottom line is suffering like hell.
So we're now becoming like a cost saving tool, and it's probably the only key driver that we need to push forward is you have to perceive us as a no install-, a very low install costs, and I will be able to have a significant impact on that expense line immediately. Well, within the first two months, right? We have to learn about it. But quickly it's going to contribute to that bottom line number, which is now the game, right?
James Dice: [00:58:20] Yeah. Yeah. I mean, I think about how like, you know, it's May 7th. So, you know, at mid March, it was all about how there's no one in the building, right? So can we shut it down completely? Now, as we start to talk about reoccupying, there's people like me that are probably going to be working from home for three more months, I would think. There's not a whole lot of reason for me to go back to the office, but there are quite a few people that are now starting to trickle back in, right? So a solution like this can help keep the energy-, you know, it's not fully shut down, but it can keep things tailored to how many people are actually there.
Jean Simone: [00:58:54] Yeah, I mean, absolutely. I mean, it's-, because of the retraining and the learning, the self-learning, it's adapting quickly to the new configuration. And, and you're right, it's going to be a moving target, right? Some people will go back, not everybody. Take a typical tower and downtown core, probably now it's either completely shut, it's basically on set back 24/7. Or there may be one or two floor because essential services are working on these floor. And then slowly you're going to have this company on another floor starting, but starting with social distancing. So probably half the employees only are going to come back and not the other half, or they're going to do rotation shift or whatever.
So, that control sequence will need to be modulated as a moving target from day to day as this confinement modulation is happening. And suddenly we could have a new research and then the confinement is back quite rapidly. So you need to rechange everything. So we're pretty convinced that the AI will bring a lot of value there. We will continuously trying to reset and basically follow that, that behavior shift, which is happening on a weekly basis. So, you know, once again, it would be feasible to do by sending a control technician to readjust the control sequence on a weekly basis. But I'm not sure how many people will really do that.
[01:00:46]
What I call the BrainBox 2.0, which is, okay, imagine if all of these AI engine, which are operating in their own building, they're not aware there's other AI engine working in a building across the street. Right? So imagine if we were connecting them together and they were kind of becoming aware that there is other AI engine deployed another building in the same downtown core. And what happened if they share information, then they start to work as a team to the benefit of the grid. So, you know, if the grid would be like sending signal and these AI engine , knowing what's happening in the next hours in term of prediction could be able to play a strategy to the benefit of the grid, especially if there's a reward allocated to them, by the utility. So we call this the swarm AI, and it's something that we will start working on pretty soon.
What did you think about these highlights? Let us know in the comments.
Note: transcript was created using an imperfect machine learning tool and lightly edited by a human (so you can get the gist). Please forgive errors!
James Dice: [00:00:00] Hello friends. Welcome to the Nexus, a smart buildings technology podcast for smart humans. I'm your host, James Dice. If we haven't met before, I write a weekly newsletter on the same topic. It's also called Nexus. Each week I share what I've learned, my opinions, and what I'm excited about in the quickly evolving world of intelligent buildings. Readers have called Nexus the best way to stay up to date on the future of this industry without all the marketing fluff. You can check it out and subscribe at nexus.substack.com or click the link in the show notes.
Since starting the Nexus newsletter, many of you have reached out to me wanting to talk shop, and we have. After a few weeks of those wonderful conversations, I realized I needed to record and share them with our growing community. So here we are. The Nexus podcast is born. This is our chance to explore and learn with the brightest in our industry together.
One more quick note before we get to this week's episode. I'm a researcher at the National Renewable Energy Laboratory, otherwise known as NREL. All opinions expressed on this podcast belong solely to me or the guest. No resources from NREL are used to support Nexus, and NREL does not endorse or support any aspect of Nexus.
Okay. Here we go. Episode 8 is a conversation with Jean-Simon Venne, the CTO and Co-Founder of BrainBox AI, an advanced supervisory control platform for autonomous buildings. We do a deep dive on BrainBox AI, including how they use artificial intelligence to save energy and to adapt building operations to occupancy patterns of the pandemic.
We talk about the problem with existing control systems and how AI helps supplement them. We talk about how AI and humans can be a great team and how the BrainBox team is approaching that teamwork. Finally, we talk about how autonomous buildings can lead to an autonomous and cleaner electric grid. And much, much more.
This episode of the podcast is directly funded by listeners like you who have joined the Nexus Pro membership community. You can find info on how to join and support the podcast at nexus.substack.com. You also find the show notes, which has links to BrainBox AI's website and Jean-Simon's LinkedIn page. Without further ado, please enjoy Nexus Podcast Episode 8.
Jean-Simon, welcome to the show.
Jean Simone: [00:02:21] Thank you. Good morning.
James Dice: [00:02:23] Yeah. So why don't you give us your personal background, and then a little bit about BrainBox.
Jean Simone: [00:02:28] Yeah, absolutely. So, I've been working in technology pretty much all my career, evolving with the technology. As you remember in the eighties, it was not the way it is today. So I kind of followed that path. I would say that my previous said adventure before BrainBox was all about optimizing energy in buildings, but I would say the old way. No offense here, but it was about, you know, big retrofit of existing equipment, let's put a new boiler, more efficient, let's try to put some economizer on existing equipment. So always big construction project, a lot of capex, time to execute a project, you know, engineering plan and specification, manage the construction, commissioning. You're easily talking 12, 18 months. And then you hope, like everything is tweaked the way it should and the saving will be, will be at the meeting point. Right. And, and usually they are, and, you are, you're all happy and you leave, you leave that project thinking job done. And that adventure was actually more on the ESCO side, so we were paid a percentage of the saving over the years.
Usually first year, all good. You know, and then the drifts start this, this infamous drift that you see happening and you can't, you can't really put your finger on exactly why it's happening, but it's happening. And it's a very slow drift. And of course, the first year, it's not really an issue because it's, it's such a small percentage that you barely notice it. But at the year two, when you have that measure and verification annual meeting, Suddenly on the year two, it's becoming more like annoying. You see that saving slowly disappearing like the snow in the spring, and you, you start to wonder, you know, gee, this, this ESCO project is going on for another four or five years. So you, you anticipate, as you know, as an engineer, you start to anticipate what this meeting will be in two, three years. And of course it's following that path, so by the time you were on year four, it's a very unpleasant meeting ,where it's always about debate and why the savings are not as good as they were the first year, et cetera. So it generates a lot of frustration.
And when you, you dive into the nature of the problem, you say, oh wow. I mean, this is thousands of small action that on a day to day operator are taking, and it's slowly de-focusing the system and it's kind of making-, trying to fix some problem, you make other problem showing up. And we should do continuous commissioning. Oh, let's do that. Actually, I went there, so let's do continuous commissioning. It's an after sales service type of thing, we're going to charge that much per month. We're going to build this control center with top gun control people, mechanical engineer. They will remotely make sure the system is always perfectly set up, and we're going to offer that service.
And then you hit that wall of, well, first of all, customer building operator or building owner, not really inclined to pay for such a service, don't understand why should I pay for that? I mean, why is it degrading? So it's, it's a very hard sell, but then you have another systemic problem, which is there's just not enough control expert or control engineer, mechanical engineer on the market available to offer that service on a massive scale. It's just not possible. Even if the customer were willing to pay for it, you would not find all of the people to offer that service for the entire building stack. So, so you're check mate before you start. And that's where the I'd start to wonder, wait a sec. I mean, can algorithm do that continuous commissioning instead of people?
And then you kind of fixed, with one stone, you fixed that double barrel problem, which is an algorithm would be costing a lot less than putting human 24/7 in a control center. They don't take breaks. They don't go on vacation. And at the same time, so the service will be very, very low cost and the same time you could scale, which was literally impossible with humans.
So that, that's how it happened in term of, that put me on that path to, maybe we should use technology to fix that problem.
James Dice: [00:06:46] Cool. Yeah. You just described like most of my career there. Yeah, I've done a lot of that old way style of energy efficiency. Yeah. From
Jean Simone: [00:06:55] Frustrating. It's always very frustrating. Yeah.
James Dice: [00:06:58] Yeah, definitely. And yeah, my most recent role was to develop a monitoring based commissioning technology stack, and then also the services to wrap around it, so I'm very familiar with everything that you just described.
So, cool. So how did, how did you then take that experience into starting BrainBox? And when was that?
Jean Simone: [00:07:19] It was in the 2016, looking at the autonomous car. I'm kind of fascinated by it. So you know, they, they kind of installed this, this, equipment on the car, which is, we look at all that equipment, it's actually costing more than the car itself. And they do that to gather or to generate the data they need, but once it's done and they have all of that data about the surrounding of the car in real time, then they basically use that data to predict that future of where all of these moving objects around the car will be in five seconds and 20 seconds and 30 seconds. And knowing with a very high accuracy what this immediate future is.
Then what is the term of operational research? What is the optimal trajectory or the safest trajectory into that future that I should follow? So then, then that become your control strategy for the car itself. Like should you go left, right, or a right, accelerate, slow down. And they're then doing that with incredible success.
I mean, it's a very complex environment. They have milliseconds to calculate and call the shots, and they're doing it. Actually there was one stats that completely flabbergasted me, it was that they were able to learn to behaviors of squirrels and cats. And if you think about it, then we all notice it, like the squirrels have this behavior as they, they're not sure if they're going to go, they don't go, they go, and then suddenly they go. And then when they realized that there's something that is not going the way they want, they stop in the middle of the street. And they question themselves again, which is the worst possible thing, right? And and as a human, we learn over time that the squirrel will probably do that. He's not going to continue his run. He's probably going to stop trying to assess a situation, and that's how I'm going to hit it. So we anticipate that. So they managed to do the same on the AI side, anticipating the squirrel behavior.
And I look at that and like, wow. If they could do this, we should be able to do the same in HVAC. We have a lot less moving part, a lot less risk. And on top of it, I have the luxury of time. I don't have millisecond to calculate and call a shot. I have minutes. Because, you know, changing the temperature in a room, takes what, 10, 20 minutes, depending on the size of that room? So I got this inertia on the thermodynamic side, which is giving me the luxury of time to calculate. So there's no reason it should not work. So that was the ticking of the idea of saying, oh, we should do that. Use the existing AI techniques, use an autonomous car, and use it for doing an autonomous HVAC. It's how it's really started around 2016.
James Dice: [00:10:05] Cool, okay. I want to dive into the autonomy piece a little bit. Let's back up though, before we get there. Why do we need autonomous buildings? Like what are the other problems with controls specifically, you know, besides the fact that buildings aren't energy efficient and they drift, like you talked about earlier, what are the issues with the actual old style of controls?
Jean Simone: [00:10:29] So, yeah, so there's the entire continuous commissioning aspect, which we already spoke about. And yes, at the beginning, that was the target. But then we realized that if we're able to predict the future of what will be happening in that building with a super high accuracy or about 99%, we basically know the future. And when we realized that the deep learning algo were, were giving us that accuracy for hours and hours ahead, so we knew that in three hours, that room will be too hot. It will hit the cooling set point in three hours, in four hours, in six hours, and we realized that we had that, that luxury to basically know the future of that building with a very, very high accuracy. You went like, wow. I mean, we could do more than continuous commissioning.
We could basically design our control strategy right now to have a better future. And it's, it's kind of going into this, this concept of, when you think about it, all of our HVAC right now, they're reactive. And even if you use a PID loop, you're reactive. So there's something which is happening now, and you're reacting to it. But you're already into that climb or in that descent, you're already going through that event. You're not into the past or looking at the future. You're really into the present. And all of our HVAC systems are reactive. I mean, they are programed which if this value trigger this action until you have this other value, then you could release or you could stop whatever action you were doing.
The entire logic of our programmation is like that, is reactive. So it's kind of similar, then, if you were driving a car and I was blocking your windshield with like a piece of wood or whatever. You're basically allowed to look on both side. So you see what's happening right now. You're allowed to look in the past. You have your rear mirror. But you don't, you're not allowed to look in the front of the car toward what's coming at you. So you'll be going through all kinds of surprise, and you're going to react to the surprise, but you're already going through that surprise. You don't anticipate.
So if you were driving like that, it would definitely cost you more energy because you were not able to anticipate, should accelerate or slow down or just let it go. And it will be very, very uncomfortable for whoever passenger is with you in that car. And I will not mention the damage that you're probably going to create on your car.
James Dice: [00:12:56] Yeah, that sounds terrifying.
Jean Simone: [00:12:58] But that, when you think about it, all of our HVAC systems are acting like that. Because we don't know the future. So we're, we're following the schedule. We're following actions that are happening already, and we're reacting to these actions by a reaction. And we're trying to stabilize as we go. So knowing the future, suddenly we could start to do preemptive action that would either improve the comfort, or reduce the quantity of kilowatt that our spent to react to an event.
And when you're at full throttle, let's say it's super hot in the summer or it's super cold in the winter, and the system are at maximum capacity. Knowing the future is not really helping you. You're already at maximum capacity. You will be running at maximum capacity for the next three hours. You're not going to save anything, but there's a lot of other periods during the year where you're not running at maximum capacity of equipment and knowing what's going to happen, and you could do preemptive action, which will save some money, and at the same time will reduce the discomfort that these actions and reactions are creating.
James Dice: [00:13:55] Yeah, like 99% of the time. Other times besides peak times.
Jean Simone: [00:14:00] Yeah. Yeah. Whatever is the building and how it was designed, right?
James Dice: [00:14:03] Yeah. Yeah, so I'm writing a piece right now for Nexus about how the old style of supervisory controls are really just quite dumb when you think about it. You have a system that requires a very expensive controller, right? And then you have a server that groups all of those very expensive controllers together. So you're talking about many, like tens of thousands of dollars at this point, and now you have this system that basically is a database that doesn't work that well, a trending device or visualization device for those trends that really doesn't really help you that much. You have some alarms that no one pays attention to. You have some schedules that are easily over-ridden and really aren't that sophisticated also. And you have some graphics that are usually wrong, right?
So it's just like this system that is really 10, 20, I don't know, maybe you could say 30-, it's like a mainframe system from before I was born. Right. So, uh, yeah. So-
Jean Simone: [00:15:04] Careful, I played on them.
James Dice: [00:15:07] Yeah. So I mean, yeah. So compare that, like I've been saying, compare that to the iPhone 11 I had just got a couple of weeks ago, and it's just night and day. You have a very dumb system versus today's smart technology.
Jean Simone: [00:15:23] Yeah. And I mean the beauty is interesting because the people already paid a lot of capex to install these controlled systems in the building, which are generating an incredible quantity of data. But that data is used on the incident of the moment, maybe kept for a few weeks in logs, but then is overwritten, right. And there is a very, very few buildings which are keeping like years of history for the granularity of all of the data point, with the very small reading interval. But the data, I mean, you already paid to generate that data, so that the big heavy lifting on the capex is done.
Think of the autonomous car. I mean equip a car to be autonomous. The big heavy lifting is all of the electronics that you need to onboard on that car. And that's the barrier to have a lot more autonomous car. I mean, who has the money to buy these type of car? The electronic onboard is more expensive than the car itself.
So, but on the building site, it's already done. People already paid huge amount of money to build these control system, which are generating a fascinating amount of data. But we're not using fully the potential of that data to create a lot more value with it.
James Dice: [00:16:34] Totally. Alright. So I want to dive into the product now. So first thing, kind of paint besides prediction, take us through kind of like a day in the life of an autonomous building.
Jean Simone: [00:16:47] So there's really the two big component once it's in operation, there's, there's the predictive. So you want to use these deep learning model neural network. Think of LSTM as an example, which gives you this very high accuracy in term of prediction, once you have enough data so that-, you know, one of the issues is you need to accumulate enough historic data to start running these systems. So easily, you know, five, six, seven weeks of data within the same season. So it's not something you could kick in within two days and say, well, ah, we have AI in this building. So you need to connect, you need to accumulate that data, and only when you have a critical mass within the same season, then you could really kick in, deploy these type of tools. So there is barrier, and you have to respect them.
So this prediction is giving you this vision of the immediate future. So on the air side, you really want to tackle the next three hours. Even though we have more, the neural network will give you a high accuracy of what's going to be happening on the air side for the next, six, seven, eight, nine, and even few days. This is how powerful it is. But on the air side control, you don't really need more than three hours ahead of you.
On the water side, think of, you know, boiler chiller water networks, then you really want more six to eight hours, because the inertia of producing water at a certain temperature is something which is prepared in a longer timeframe than, than the air side, makeup air if you want. So that prediction is the first fundamental part.
So doing prediction, on the AI, we have a saying which says, doing the prediction is the easy part. Doing the control is the hard part. Because the big issue we have, like with building, when you think about it, is you're not trying to resolve one linear problem. So let's say one linear problem could be, I want to maintain the temperature all the ways within my dead band of the set point. So that would be like a linear problem and you would not need neural network to do that. A classical machine learning tool would do it. But then you want to do, okay, I want to do that, but at the same time, I would like to do it with the minimum quantity of kilowatt hour that I'm spending over time. Oh, but then you have a second problem that you're trying to resolve and you're trying to optimize both of these problems at the same time.
But then, Oh, wait a sec. This is not it. Because the utility, they're charging us on the power factor too. So even though I am saving kilowatt hour, if I'm creating a super peak on the power side, I'm not going to win at the end of the month when I'm going to receive that bill. Actually, that could create a worst case, because the power factor of the bill, it could be very intensive in dollars. So we want also to optimize that kilowatt in power, as a third line that you want to optimize at the same time. So, Oh, this is great. So then we're starting to have a more complex problem that you want to resolve. So you want to optimize these three lines in parallel.
But wait a sec. We don't want to create some cycling on the equipment. Right? Because we will be slowly destroying too fast like a pump or a ventilator. So, so I want to make sure I'm not cycling anything here. I'm treating his equipment with respect and I'm not creating equipment problem down the road. So, Oh, that's a fourth line that I need to optimize in parallel to the first three.
And then your problem is becoming so complex. It would be resolvable. I mean, we would put like probably, you know, you, me and a couple of other engineer, a couple of stats model around a table. We're going to crunch this, maybe with MATLAB, and we would come with the solution. Probably going to take us a few hours. But we would come with that solution of what is the optimal control for the next hour. By the time we do it, of course, that hour is long gone. So whatever we came up as a solution is too late, because we should have done that in half a second if you want to be efficient. And just imagine the cost of all of us around that table. And we would have, of course, to work 24/7, day and night because you know, we need to operate that building and that's only one building.
So it's, it's all feasible by hand, but the power of like a deep, reinforcement learning. And that's the same, the same thinking than the video game. So when the AI is playing against us on a video game, it is quickly understanding what are the rewards, so how to win the game. So if I want to win the game, I need to maximize basically my points and to get points, there are several ways you could get points, right? So quickly the AI is understanding how to get these reward, and how to score maximum value and win the game.
So it's the same, same technique we're using, but we're kind of presenting to the AI as a game, saying we have all of these paths and you need to optimize all of them in parallel. And if you manage to do like a global balance, you're going to win the game. And that's what we call the deep reinforcement learning. And this is what we're applying. And it's producing a better control strategy than the typical control sequence, which was already there in the building. Which is, as you described earlier, you know, based on fixed schedule, which are not always tweaked the right way, which is based on reaction. So it's reacting. Remember, the AI knows what's coming, so he's got a clear advantage on that control sequence because he knows what's going to be the demand for let's say for the chiller in two to three hours. So it could pre-produce water without creating a super peak.
There was a good example that I like to bring forward in a very high rise, a very big high rise in Montreal. One of the issues they had is that their chiller is working at full, full throttle. A Beautiful day outside in the middle of July, very hot, and it's sunny, and then suddenly these clouds move in around 2:00 PM, very thick cloud, and suddenly it become very dark.
So within 15 minutes, there's hundreds of tons of chilled water that are over capacity in that building. So they start to shut down the systems and lower the production because they're in full excess capacity. But the AI knows these clouds are coming, because right now, when you're taking this pilot weather data, and you know exactly the movement of the cloud and the cells, and the thickness of the cloud, and you know where are they going to be. And so we could anticipate that you're going to have like 200 tons of chilled water in excess by 2:00 PM this afternoon. And so you could start to glide it in advance, and slowly lower your production on the cold water in anticipation of that, and not reacting to an event which is happening suddenly.
So that's just an example, but it gives you a very smooth curve, knowing the future and it's saving kilowatt here and there, all day long.
James Dice: [00:23:43] Totally. That's a great example. So I want to circle back real quick, so comparing autonomous buildings to autonomous cars, it seems like it's a different output problem. So with a car, it's like I am either going to brake, I'm going to accelerate, I'm gonna turn left, I'm gonna turn right. There's, there's just limited output, right? Versus a building, it seems like , like you said, we're optimizing for more variables. Is that the right way to think about it?
Jean Simone: [00:24:08] Yeah, I would say that you're right, there's more variables on the HVAC than on the autonomous car, but we have less moving part. So, so you're right, I mean the autonomous car has less variables that they have to treat, but they have so many moving objects around them that it's making the problem very complex.
On our side, we have very few moving object, it's actually a very static environment, but we have a lot more variable to consider to figure out what is the thermodynamic equation that will be happening here. And then again, I mean, the AI does not, artificial intelligence does not know the thermodynamic handbook that we all study in engineering, because it's only data driven. But it's figuring out the: if this happened, this will happen. It's figuring all of these action reaction that you see typically in the building, and it's becoming super customized on one building. So once it's been learning on one building, it's what we call completely burn.
That model cannot be taken and used in another building because each building is so different on the way it's behaving in term of the thermodynamic side that, that really, once it's learned that building it, yeah, it's only good for that building.
James Dice: [00:25:18] Okay. And you also mentioned this acronym LSTM. What's that stand for?
Jean Simone: [00:25:23] Yeah, that's one of the-, there's a lot of different neural networks, that are out there that could be used as one of the models. We're using several types. One of the things we do is, is we basically when we do the learning part is we test different neural network model and we figure out which one is the best fit for this specific building.
James Dice: [00:25:41] Hmm. Got it. Okay, cool. So tell us about kind of, the rest of the stack. So I'm assuming all those AI algorithms are in the cloud, right? So how does the, how do you get the product? I'm assuming there's a box, right? So it's called BrainBox. I'm assuming there's a box. So, tell us about the entire stack from the building up to the cloud.
Jean Simone: [00:26:03] Yeah. So one of the big challenges we had to resolve-, and it was nothing to do with artificial intelligence, was the fact that there is a lot of different controllers out there on the HVAC side. We actually did a count and we went all the way up to about 700 HVAC controlled protocol in the world. Some of them are, you know, were created by company that don't exist anymore. They're not being supported anymore. Some of them were bought by a bigger company and they're trying to support the line and at the same time convince a customer to, to switch over. There's also a lot of different versions, so, so on each of these control protocol there is different version, and a building may necessarily upgrade their control systems. They try to, you know, go as long as they can without upgrading because they want to avoid the expense.
So it's a pure nightmare. You want to interface with that kind of ecosystem, and some of these protocols are completely closed. I mean, there's absolutely no way you could understand that language. There's no documentation. The company does not exist anymore. Good luck to find somebody that knows about it. So we had to create an edge device, a physical edge device. That we could install in the building, and would be able to connect to that existing control system and talk the same language.
So of course, you know, BACnet was the easy one. But think about LonWorks. Think about Modbus. Honeywell is an interesting protocol just to name one. So it's, you need to be able to connect that box and that box is able to talk that language to first of all, discover what are the points, so an auto discovery. Then read the point. And when we read the point for the first time, let me just tell you that it's a, it's a very rare exception when we hit the building, which is already Haystack tag. Most of the building, it's a very weird nomenclature, which is sometime created by the technician that did that set up. We see building, which is, let's say, half in English, half in Spanish. And we see building where people are giving names to the equipment. I mean so instead of calling a fan a fan, they give it the name of a person. You know, this is Andy and it's a fan, but they don't say it's a fan, just this is Andy. So we have to figure out that Andy is a fan. So, you know, nomenclature is a very creative world in the building HVAC control. And so that's a step that, that we need to do manually.
And actually we're working with the NREL in Colorado, and we are creating what we call the Autobot. So the Autobot is a piece of AI which is doing that mapping, that conversion of whatever name they were using in that building to a standard, Haystack tag nomenclature, because it's only when you organize your data in that fashion that the artificial intelligence could understand what it is.
And that is the step that we call mapping, so still manual process in the motion to becoming AI too in terms of the mapping. So if the AI could do about 80% of that mapping, and we still have human doing 20% of the mapping, I think that would be very, very happy if we reached that level of applying the AI for the mapping, naming conversion.
So once that step is done, this edge device connected to the existing controller of the building will start to read the data and send it to the cloud. So we have different way to do that, in term of secure connection. And that reading goes in the cloud where it accumulate in a database, which is specific for that building.
And then we have to wait. We basically have to wait that period of time to accumulate enough data. We're also in the cloud mixing that data coming from the building with detailed, detailed whether data that I mentioned previously. So you want that, that special weather data, which is giving you the wind, the wind direction, the wind gust, the cloud thickness. So we're not just talking about humidity and temperature here. We're really talking about detailed weather, because there's a lot of correlation between what's happening inside the building with sometime driver, like the wind direction and/or cloud thickness, because that's giving you directly the solar radiation intensity, that cloud thickness.
So that weather data is accumulated in parallel on the same timeline than the data points were taken for the building. During that period of time that I mentioned, you know, five, six, seven, eight weeks, in the same season. And it's only then that we have enough data that you could start to apply this prediction.
And for us, this prediction step is very important because it's giving us the quality control that, yes, we could start to do automatic control because the prediction is good. And it's also something you want to keep doing in terms of prediction analysis. Because when you see a degration of the prediction, it's your signal that you should also retrain your AI.
So that's happening in different situations. It could be happening when there's a season change. So you were, you were training yourself, you were operating during the wintertime, and then you were getting in the spring, you're going to start to see new behavior happening in terms of the weather, of course. And that will have a different impact. Switching from heating to cooling is an interesting aspect, and it's happening in the reverse order in the fall, that requires a retraining of your neural network, so they discover new behavior that's happening all along during the first year.
But it also could be a user behavior change. So a tenant is leaving the 10th floor. He's going to another the building. So suddenly the 10th floor is empty, and then then they will do construction on the 10th floor for the new tenant that's moving in. And then there's going to be a new tenant. Right now we are seeing in the COVID-19 crisis, all kinds of behavior change on a tenant side. So there's a tower, which on 30 floors, there's only two floor, which are still occupied, and it's occupied by a government department, which used to work nine to five, now they're in crisis management, so they're working from six in the morning to midnight. So and all of the other floors are now empty. So who's going there to change the entire control sequence to adapt to this new reality of that tower, right? Feasible. Is it being done? Well, the AI automatically recognized there's a big shift happening and that night will retrain itself to this new behavior that it's understanding. And after a few days, we'll be now completely understanding the new setup of the tower and we'll be trying to optimize that new set up.
Well, that retraining is happening in the cloud. So most of the AI is happening in the cloud. And then we have a lot of what we call control algorithm. So we have about 20 something of these algorithm , and they all focus on one part of the system. Either they focus on the entire tower, like we have an algorithm which is optimizing the schedule. So what time should we start the system this morning? What time can we stop the system in the evening if there's a schedule, if it's that type of building. We have other algorithms which are trying to optimize the temperature of the water, of the hot water being supplied into the system at any given time. So same thing on the chiller side. We have other algorithm which are focusing on power peaks. So how can we manage to not create another peak of power that month in the tower.
We have algorithm which are focusing only on air handling unit. So when is the optimal time to start the cooling stage one or cooling stage two? We have other algorithm working only on fan speed if you have variable speed drive. So what is right now my optimal static pressure in that duct at any given time.
And we have to make all of these algorithms work together. So, we need a coach, which is making sure that the team is doing a good teamwork and not working one against the other. So that's the complexity of it. So most of these tactical algorithms are located at the edge device, because the ones that are doing these decision on real time are located at the edge device, but they work in tandem with the crunching happening on the cloud side.
James Dice: [00:33:58] Yeah. So all the heavy lifting is happening in the cloud.
Jean Simone: [00:34:01] Yeah, because of the processing power that we need.
James Dice: [00:34:04] Yeah. I think what you just described was one the answers to one of my follow up questions, which was, if you lose internet, that local edge device has basically the latest algorithms and/or control sequences in our industry's terminology. It has the latest and greatest from the cloud. And it's getting those every couple minutes as an update, or how does that work?
Jean Simone: [00:34:26] Yeah, we try to standardize. On the air side we are, we're standardizing around five minutes. So we're reading the point every five minutes, we're calculating that prediction every five minutes, and then we're deciding what is the optimal control strategy every five minutes. And then we're waiting another five minutes through to reassess the situation.
We decide to work at the zone level. So our fundamental calculation block is the zone level. And when we want to know what is my strategy for the air handling unit, which let's say might be serving six zones, we basically aggregate the prediction and a control strategy for all six zones together. And then you have your control strategy and your prediction for that air handling unit. And we keep backing it up. If you want to know what's going to be your chiller strategy, well you back up from that valve, which is opening the cooling for that air handing unit, you're backing it up to aggregate all of your air handling units in the building to know basically, what is your best strategy for your chiller at any given time.
So we figure that structuring it like this is giving us the ability that really we don't care the size of the building. So it could be a retail with two zones, or you know, two rooftop, two zone, a small retail store. We're managing them, the two, and it's pretty simple. But a big high rise would be just like maybe 300, 400 zone, aggregateed together like Lego block. So on our side it's just more zone and it's not more complicated. It just takes more time to onboard or to configure it. So that's how we kind of build it.
On the water side, that five minute cycle is too long. So you want to have more like a minute cycle, because that latency of five minutes could become a problem on the water side. So you really want to, you want to more have like a minute cycle when you're playing on the liquid side.
James Dice: [00:36:22] Okay. And what does happen if you lose connection with the cloud?
Jean Simone: [00:36:26] Yeah, sorry, I forgot the main, the main question. Right. So, yeah, so as like autonomous car, the safeties are more important than the action , than your control strategy. Because the last thing you want is to have artificial intelligence which is starting to take the wrong decision and basically controlling your building, right? So, so you need to put a lot of safety in place to monitor if the AI is doing-, first of all, is the AI working? Are we extracting the data? Are we getting the data? Do we have that communication with the cloud? Is the quality of the decision of the algorithm the right quality?
So you need to put basically other algorithm that are checking if the quality that the decision being taken is the appropriate quality, because you want to detect, through all of these safety, if you have a problem. And if you have a problem, you want to de-engage the AI. So one of the problem could be, as you mentioned, that we lose the communication with the cloud. So we need to detect that, and we need to start to monitor that. And basically if the communication is not coming back within a few seconds, you want to start to de-engage.
So de-engage means that you're starting to revert all of your action in a very slow fashion. So we want to have a slow, smooth landing, and you want to basically give back the control to the existing control sequence, which is still there in the background. It's still doing its job, and you want to basically put it back in control of the control sequence. So you're basically de-engaging all of our action and reverting back to that control sequence. And once the communication is reestablished, and even then you want to make sure that it's stable to avoid cycling on the communication problem, then you want to start to reengage automatically.
So this safety, you will find that in an autonomous car, it is actually more safety, about how you're gonna drive that car than there is control strategy of the car, so we have exactly the same situation where we have more safety, especially on the edge device and on the cloud than we have scripts doing the control itself.
James Dice: [00:38:39] Hmm. Yeah, okay. So it's basically when you have full communication you're able to override all of the controllers and send set points and commands to them, basically. And then when you don't have communication, you're then just removing that priority over the set points they already have, basically. Okay, cool.
Jean Simone: [00:38:59] Yeah, yeah.
James Dice: [00:38:59] So a lot of different ways I want to go with that. That's a fascinating view at your stack and how all the things work. So thank you.
So one of the things that strikes me is when I try to-, so shout out to Maddie. Maddie is the one that edits the podcast and helps me produce these. And one of the things that her and I talk about is like, what if I'm explaining this to someone who's new to the industry? And a lot of times it's very difficult. That's why you're laughing so-.
Jean Simone: [00:39:26] Oh, yeah.
James Dice: [00:39:26] So I think what you just described, while it's super high tech, it's, for someone that's just out of school, right, it's not that high tech. Right? It seems like, like the question that comes to mind for someone that's new is like, why weren't buildings doing this already? And one of the things that I like to talk about is all of the enabling technologies that enabled us to, or you guys to build this impressive stack of technologies.
And so I think, I don't want to gloss over the fact that all these are actually pretty new, for buildings at least, right? So you talked about the edge device, which I know you guys use Raspberry Pi's. You talked about BACnet. I mean, that's not that new, but it's also, it is kind of new, right? It's new in terms of the-, the control contractors and OEMs are now being forced by owners to use BACnet or other open protocols. And that's relatively new, right? You guys use cell modems, and so that whole infrastructure you're building on top of that. All of these AI algorithms, so that's not something I'm familiar with, but when did those start coming on the map, in the advent of all technologies in the cloud and that sort of thing?
Jean Simone: [00:40:35] Yeah. Before going on the AI side, I might just back up because you're absolutely right. I mean, just think of storage, data storage. I mean, 10 year ago, 15 year ago, it was so expensive. That if I would have come to any CFO and say, you know what, you have a tower, there's about 40,000 point of HVAC data, I'm going to be collecting a reading every five minutes of that, and I'm going to accumulate it, and then I'm going to use that, you know. The cost of storage, incremental because I keep everything, you know, you want to keep everything when you do AI. We're probably going to have shut down the project right there, because the CFO would have said, are you crazy? It's way too expensive storage of data. But today, cost of storage of data is becoming so low that actually, we're collecting all of the points of the building even though there is probably a few thousand point that we're pretty convinced that we will never need, we're keeping them anyway, because we might discover in two, three years that finally we did find a usage for that data point, and we're very happy to have collected since the beginning. But I mean, the price of storage right now, it's really not an issue.
You mentioned the cellular connection, yes. We're using an IoT SIM card from, from AT&T and Bell Canada. The cost per month is now becoming low enough that, you know what, let's do it. But not long ago, that's cell connection monthly recurring costs plus transmission costs, right, would have probably killed the project right there. Not viable.
So, and then you get into the CPU, I mean the GPU, the capacity to crunch data. I mean, I'm telling my kids that, you know, their cell phone is more powerful than the most powerful computer we had at my university in the eighties. And I showed them a picture of that computer. It had a name, and we were allowed probably 15 minutes on it per week. And you had to justify that it was a worthwhile calculation, that you really need access to it. So when you think about that, I mean. It would not have been possible before to do everything we're doing right now.
So, yeah, I mean, a neural network is a new thing. I mean, a lot of people are talking on the AI about, you know, but it is, it's a new thing. I mean, it's just really flourished in the last six, seven years. Right? Before that, it was not on the radar. I mean, it was being talked about since the eighties. And you, when you think of the, Yoshua Bengio, Lacan, I mean, they were pushing for that for the last 30 years, but nobody was listening to them. Actually they were told that, you know, they were wrong many times. And it's only in the last decade that they had their glory moment where everybody turned back to them and said, wow, you were right. This is the way to go. This is how we're going to do deep learning, deep reinforcement learning, and it's been exploding in the last decade now.
And there's so many possible applications that, I mean, we're just starting to see a real application hitting the market, which are bringing a lot of value, and it's creating new business model, either for startup or for large existing company, but it's just starting to basically shape a new business model, and bringing us into a new world, in term of what can we do better. So, we're really in that stream of applying neural network and trying to create new business model with it.
James Dice: [00:44:11] Fascinating. Yeah. I love that story of like, why now? And it just, all of those enabling technologies coming together, it's pretty awesome.
So, okay, let's talk about, challenges. So, and I've been in a lot of buildings have done a lot of projects, like you talked about earlier, the kind of the old way. And one of the things that, when I think about if I were to go, you know, knock on someone's door and, and sell this technology, a couple of challenges come to mind. So, what about buildings that sort of are in physically poor shape? So, you guys have pulled a lot of data, but what if I can't trust that data? What if sensors need to be calibrated? And what if valves are stuck? What if dampers are stuck? How do you guys approach that, non-software side of things?
Jean Simone: [00:44:57] Yeah. So I mean, and it's the case in every building, right? I mean, it's the reality we're in. So we're, we're limited by the state of the building on the control side. So, if the data is, not calibrated right, you know, we, we see these things, you know, this room is now at 600 fahrenheit. So we have to discard all of these data points, and we're basically blind. So when we're blind in some area of the building, we basically do not engage any type of control in that section of the building. And we, of course, we will notify the building operator, the building manager, the building owner, that, you know, I mean, are you aware that this is a list of points which are either not working anymore or decalibrated. These are equipment which are not working anymore. You mentioned the damper, that's very frequent. You're sending the signal to the activator to open the damper, but the damper is not moving. It's just completely jammed. So I mean, we provide this to the building operator, owner, and suggesting that maybe they should call their control contractor and get them repaired, right? But then it's up to the customer to decide if they want to do it or not. In the meantime, until it's fixed and if it gets fixed, you know, that could take times, we basically removed these area, and we just control the area that we have valid and good data, a reading. So it will create kind of the Swiss cheese, where we're controlling maybe, some of the floor, but not all of the floor, some of the area, but not all the area. And we do as much as we can with the input we get. We can not change, we cannot fix that problem. It's really a building owner, building operator decision at that point.
James Dice: [00:46:39] Yeah. I think it's just like any other thing, any other thing in buildings, is it requires not just the technology, but also the humans and the processes to maintain and maximize or optimize the service itself. So I don't think you guys are going out there and saying, you know, this is a panacea, right? It's, it's not, it's not going to fix everything on its own in its own silo. It's just like everything else that needs to be part of, an ongoing strategy for the building owner.
Jean Simone: [00:47:07] Yeah, we actually try to say, you know, there's no magic here. It's pure maths. So it's garbage in, garbage out. I mean, we're not gonna fix these type of problems.
James Dice: [00:47:16] Totally cool. Yeah, and I think, one of the things that I feel like our audience here, people that are listening to this can get out of this, is that, it's not replacing the other types of building optimization that are out there. Right? So if you think about fault detection, this plays hand in hand with fault detection in that, I mean, maybe these are better control sequence, so there'll be less control sequence related faults, ideally, right? But the physical faults will help-, like when the physical faults are fixed, the control sequences will be able to control more of the building and it'll be able to do better, a better job. So, cool.
So another type of challenge that I thought of as I think about my average building operator, so there's two challenges and they're both pride related. So, when I think about a building operator, I think about one that either has pride in his control sequences or one that has pride in the way that he operates the building without his control sequences, right? So, in the former, it's like, we have, the state of the art control sequences. Our mechanical engineer did a great job. I don't really understand them. I just know they're working really well. And the latter, it's I don't trust control sequences. I operate my building. Right? And how do you guys approach those two separate pride challenges?
Jean Simone: [00:48:37] Yeah, I mean, you're touching probably the most difficult part of any project, right. How do you make the AI work with humans? So, I don't know if you had the pleasure to sit in an autonomous car level four, but, I was told that, I was really not a good person to do that test, because I'm fighting with the AI. I want to control the wheels. I don't trust it, and it's because I've been conditioned with, you know, a lot of years of driving myself. And for me, you put me behind a wheel and I will drive, right? I'm not gonna just cross my arm and look at the show, right?
So, I think anybody, which has, you know, been maintaining a building for, for years and years, and it's becoming your building and you're very proud of what you do, and you should be, it's not an easy job. You know, I think of all these tenants which keeps complaining and you know, issues, and you're trying to maintain that level of happiness, with all of the users of the building, it's an art. So with the tool that you were given, you pushed them and you configured them through a level of which you're quite satisfied. And you did put so many hours to make it, tweak it so it does the best it could with whatever you have. So, and then suddenly you have that piece of AI coming in and saying, Oh no, we're going to go get like a lot more value out of this building. And it could be very frustrating, to see that, wait a sec. You know, do you know how many hours and years that I put into this?
So, and there's an interesting story about this. I was talking to one of these person that we're describing here, which spent his entire career in one tower. And he's about five years away from retirement, and he's the ultimate expert of that building. And he was looking at me, he said, there was nothing the AI could do better than me. And I was like, okay, yeah, interesting discussion. And I said, well, are you going home at night? And he said, yeah, yeah. Yep. Seven to three, sometimes 3:30. Okay. Perfect. So who's-, if you're not here, what happens? Because there's a shopping mall at that bottom of that tower, during the evening, you know, Thursday, Friday night, that's open until nine. You know, what happens in the shopping mall?
And he said, well, if there's a problem, they're going to call me. Okay. And you like that, to be called? No, no. No. Okay. And what happens Saturday? Shopping mall open, right? Yeah. Yep. They will call me if there's a problem. And what happens if there's something happening during the night and then? Oh, I mean, that's not good. So what about the AI is becoming your second in command, and it's managing the ship 24/7, so you could have a better night to sleep, so you, you could go on vacation two weeks somewhere in Europe. You could go for lunch, not being stressed that your pager is going to ring. And he looked at me and said, okay, now I understand.
And he looked at me and said, and maybe if I show the AI a lot of things in the next five years, maybe I could retire with peace of mind. And I said, yeah.
James Dice: [00:51:26] Yeah. And the building owner would have peace of mind as well.
Jean Simone: [00:51:29] Yeah, so, the AI is not here to replace people. AI Is here to give us the ability to do more on our daily basis. It's increasing our capability or capacity to do a lot more and to maybe focus our creativity on the most important things and letting all of the basic stuff being done by the AI, freeing you up some time to do more important things with your expertise.
James Dice: [00:51:54] Yeah. Fascinating. I don't think I'd be a good passenger in an autonomous vehicle either. I'm an awful passenger as it is. So, cool.
Well as we kind of get to wrapping up here, I want to switch over to the business side. So this is obviously a technology that's on the market. So how are you guys taking it to market? And what's your strategy there?
Jean Simone: [00:52:16] We're focusing on the first wave of building that we test AI at first. So you're talking about retail, you're talking about office tower. So these type of building are, you know, more our focus right now. But we're also stretching into other type of building now, think of data center, airport, hotels, are now becoming a new candidate for the application of this AI. The model that we-, we were trying to basically make it as easy as possible as an offering. So we want it to be a SaaS model, so it's a, it's a monthly fee that you're paying as a service. It's a very, very, very low install cost, so there's basically no downside to go and implement this AI. So you get the box installed, you start to get the training of the AI, and then you pay a monthly fee, which is based on the square feet and the energy intensity you have in your region. So it's a fixed amount of money per square feet you're paying, which is a much lower number than the savings and the AI is going to bring to the table during the same months.
So we want it to be like a very compelling offer where there's really no risk for the building operator or building owner. They will be saving per square feet a lot more than what they're paying on a monthly fee. And the same time is, you know, you don't like it, it's like your cable service at home. You don't like it, well, terminate the service and send us back the box at the end of the month.
So very different than my previous world where we were coming in with an ESCO contract, which is we're going to go in and we're going to basically go to a wedding together where, where, you know, this contract is for the next five, eight years, and it's a very complex contract, and once you sign, you're in for the adventure and there's no backing up. Right? So I mean, we really don't want to offer something like that. So let's make something very, very easy to accept and easy to, basically say yes, value-based, really only value-based. You get the value, you pay. You don't get the value, you don't pay. And the proof is in the pudding. Right?
James Dice: [00:54:26] Hmm. Yeah. So I mean, to kind of summarize, it's cashflow positive, no capital expenses, essentially.
Jean Simone: [00:54:33] Yeah.
James Dice: [00:54:34] Cool. Okay. How about partnering with other vendors? Are you guys going direct to market, or is it channel partners, or how are you guys getting to building owners?
Jean Simone: [00:54:43] Doing a mix of different approach. So yes, directly to the market, have our own sales force, knocking on door, and offering the service to customer. But a big, big, also emphasis on channel, so want to have like the integrator, so the SI in a different region, being able to resell this service. And also having another layer, which is the OEM, so having. control manufacturer embedding this directly as a container or a driver into their controller. So it's AI BrianBox ready. And as that controller is installed, the service is ready to be offered if they want to subscribe to it, but it's already embedded into the controller.
I mean, in a similar type of channel approach, we're finishing up our tridium driver. So, you have a tridium JACE, you just download our driver from the marketplace on your JACE, AX N4 compatible, and voila. I mean, you don't need to install a box in your building. You're ready to go. That JACE with our driver will do our different proximity function and connect to our cloud automatically.
James Dice: [00:55:53] Yeah, and that's one of the things that I think is so exciting here is that, you know, especially coming from my history, which is very similar to your history, is besides like the physical and pride obstacles, there isn't a whole lot of financial obstacle or, really a technology obstacle to getting this hooked up if you think about all the JACEs and, you know, even even in the future with like people white labeling this or installing it on their own supervisory controllers.
Jean Simone: [00:56:21] Yeah, we're focusing on the AI side. For us, our value we're bringing is the AI. All of the rest is the plumbing that is needed to connect and make the data accessible. So, you know, we're very open to any type of a combination that could make sense on the business side.
James Dice: [00:56:36] Cool. Okay. So, you know, it's, May 7th today. And I've been saying that on the last couple of podcasts because, we're obviously in the middle of a very tough time for our industry and for our economy as a whole. How are you seeing that changing buildings and how is it hanging your guys' approach right now?
Jean Simone: [00:56:56] It's interesting because we were talking about that a few weeks back, trying to understand what will be this new world. Last year we were selling more on we will improve the comfort, so, you know, we will reduce issues you have on the comfort side. And a lot of customers were like, it was their key priority. Yes, I want to save money, but I have this tenant on the sixth floor, which keep complaining and we can't stabilize the temperature. And if your AI could fix that, it's worth the money. And there was also a lot of, let's save the planet. If we reduce the energy intensity, I'm going to be able to say that my building consumed less per square. I'm helping the planet. That was really the spin on the sales side.
What we're seeing now is that kind of completely shift now. Now people are, okay, I need to save money. I lost two tenants. They're not paying their bill anymore. My revenue are coming down, so now I need to do whatever I can on the expense line to reduce that one, because the bottom line is suffering like hell.
So we're now becoming like a cost saving tool, and it's probably the only key driver that we need to push forward is you have to perceive us as a no install-, a very low install costs, and I will be able to have a significant impact on that expense line immediately. Well, within the first two months, right? We have to learn about it. But quickly it's going to contribute to that bottom line number, which is now the game, right?
James Dice: [00:58:20] Yeah. Yeah. I mean, I think about how like, you know, it's May 7th. So, you know, at mid March, it was all about how there's no one in the building, right? So can we shut it down completely? Now, as we start to talk about reoccupying, there's people like me that are probably going to be working from home for three more months, I would think. There's not a whole lot of reason for me to go back to the office, but there are quite a few people that are now starting to trickle back in, right? So a solution like this can help keep the energy-, you know, it's not fully shut down, but it can keep things tailored to how many people are actually there.
Jean Simone: [00:58:54] Yeah, I mean, absolutely. I mean, it's-, because of the retraining and the learning, the self-learning, it's adapting quickly to the new configuration. And, and you're right, it's going to be a moving target, right? Some people will go back, not everybody. Take a typical tower and downtown core, probably now it's either completely shut, it's basically on set back 24/7. Or there may be one or two floor because essential services are working on these floor. And then slowly you're going to have this company on another floor starting, but starting with social distancing. So probably half the employees only are going to come back and not the other half, or they're going to do rotation shift or whatever.
So, that control sequence will need to be modulated as a moving target from day to day as this confinement modulation is happening. And suddenly we could have a new research and then the confinement is back quite rapidly. So you need to rechange everything. So we're pretty convinced that the AI will bring a lot of value there. We will continuously trying to reset and basically follow that, that behavior shift, which is happening on a weekly basis. So, you know, once again, it would be feasible to do by sending a control technician to readjust the control sequence on a weekly basis. But I'm not sure how many people will really do that.
James Dice: [01:00:08] Right. Yeah. And it seems like, I mean, a lot of what we're talking about on the HVAC side is it's modulation, right? So, and I think solutions like this allow a building operator to say, how can I also modulate by operating expenses, in accordance with my revenue, which it seems like revenue is uncertain right now. So being able to tailor your expenses as much as possible is, is huge.
Alright, so I also saw you guys just had a round of fundraising, so congratulations there. What's the roadmap look like for you guys over the next year, three years, five years? Where are you guys headed?
Jean Simone: [01:00:46] Well, we're, we're selling the solution. We're already doing several installs on the international side with a few building in Australia and Ireland. We're now going into Southeast Asia. It's quite interesting because even without like the tridium driver, I mean, it's just a box that we could ship by FedEx or UPS, and we're just helping them to install it with a regular FaceTime on our phone. So it's that easy. So it's giving us the ability to, if we have a partner in Bangkok, to just ship them the box. They will install in their own customer in Thailand and, and we do everything remotely. So for us it's quite easy as long as FedEx and UPS are working, we're good. We could ship. So, it's giving us the ability in this confined environment where we could still sell and we could, we could still propogate that, that value generation and different customers throughout the world. So we're definitely going to keep pushing that, keep pushing the development of the things like the driver for tridium that I mentioned.
And then go to, what I call the BrainBox 2.0, which is, okay, imagine if all of these AI engine, which are operating in their own building, they're not aware there's other AI engine working in a building across the street. Right? So imagine if we were connecting them together and they were kind of becoming aware that there is other AI engine deployed another building in the same downtown core. And what happened if they share information, then they start to work as a team to the benefit of the grid. So, you know, if the grid would be like sending signal and these AI engine , knowing what's happening in the next hours in term of prediction could be able to play a strategy to the benefit of the grid, especially if there's a reward allocated to them, by the utility. So we call this the swarm AI, and it's something that we will start working on pretty soon.
James Dice: [01:02:42] Cool. Yeah. It's similar to something I just saw from Google where they're shifting their data center loads across the globe based on time of day, based on how clean the local grids are. So you can think about strategies like that, which are super exciting for the future. Cool.
Well, is there anything else, that we didn't cover that you wanted to sort of say to the folks?
Jean Simone: [01:03:04] No, I think we recovered a lot, and I hope it was not too, too deep, too complicated to understand. Thank you very much for the opportunity.
James Dice: [01:03:12] Yeah, absolutely. I don't think it was too deep for our group. We got, we got some smarty pants that are listening here. So and I'll put my take in the show notes and, again, thank you for, for coming on the show.
Jean Simone: [01:03:25] Hey, it was a pleasure. Thank you very much.
James Dice: [01:03:27] Alright, friends. Thanks for listening to this episode of the Nexus podcast. For more episodes like this and to get the weekly Nexus newsletter, please subscribe at nexus.substack.com. You can find the show notes of this conversation there as well. As always, please reach out on LinkedIn with any thoughts on this episode.
I'd love to hear from you. Have a great day.
Happy Thursday!
Welcome to this week’s deep dive exclusively for Nexus Pro members. It’s an honor to have you here. This deep dive is a follow up to my recent conversation with Jean-Simon Venne, CTO of BrainBox AI. I learned a lot from this conversation and want to share my takeaways and the full transcript with you below.
In case you missed it in your inbox, you can find the audio or video here:
Nexus site | Apple Podcasts | Spotify | YouTube | Add to other podcast apps
Enjoy!
—James
Disclaimer: James is a researcher at the National Renewable Energy Laboratory (NREL). All opinions expressed via Nexus emails, podcasts, or the website belong solely to James. No resources from NREL are used to support Nexus. NREL does not endorse or support any aspect of Nexus.
Well, first, let’s just agree that BrainBox AI is a slick startup.
They have the AI expertise, great marketing, a solid business model, new venture funding, a solid sales force, and leaders like Jean-Simon. I especially love the sales philosophy: they’re like a cable provider without a long term contract. You’re not getting married and can cut the cable at any time.
They’re also in on the all-star game approach… AI is their position and special sauce. They’re fine with that living on someone else’s hardware and being white-labeled and don’t need to sell the whole stack. Watch out for this crew!
My main negative reaction: I’m skeptical that this solution can be implemented in a silo, and that seems to be the message I’m getting, not from Jean-Simon, but from others at Brainbox AI. If the building loses connection with the cloud or if the data is erroneous or if the valve is physically stuck, the AI stops working on that section of the portfolio. While it might be tempting to try and package this up as a stand-alone solution, it’s only as effective as the rest of the system. So I’m worried that their sales messages could mislead building owners into thinking they just turn on AI and the building is now as efficient as it can be.
Another thought: there seems to be quite differing opinions on whether cloud-based supervisory control is feasible for buildings. Terry Herr, our next podcast guest, says it won’t work. Same with Troy Harvey. The GSA demonstrations of BuildingIQ had huge issues with it. Jean-Simon doesn’t seem to be worried about it. Soon I’ll be partnering with another engineer to share a new paradigm for supervisory controls—he’s not too worried about it either.
Finally, am I the only one that’s curious about the amount of energy this sort of AI requires in the cloud? It doesn’t do any good to save energy in the building by eating up an equivalent amount of computing power in the cloud. How do those two numbers compare?
What do you think?
[00:10:29]
We realized that if we're able to predict the future of what will be happening in that building with a super high accuracy or about 99%, we basically know the future. And when we realized that the deep learning algo were, were giving us that accuracy for hours and hours ahead, so we knew that in three hours, that room will be too hot. It will hit the cooling set point in three hours, in four hours, in six hours, and we realized that we had that, that luxury to basically know the future of that building with a very, very high accuracy. You went like, wow. I mean, we could do more than continuous commissioning.
We could basically design our control strategy right now to have a better future. And it's, it's kind of going into this, this concept of, when you think about it, all of our HVAC right now, they're reactive. And even if you use a PID loop, you're reactive. So there's something which is happening now, and you're reacting to it. But you're already into that climb or in that descent, you're already going through that event. You're not into the past or looking at the future. You're really into the present. And all of our HVAC systems are reactive. I mean, they are programed which if this value trigger this action until you have this other value, then you could release or you could stop whatever action you were doing.
The entire logic of our programmation is like that, is reactive. So it's kind of similar, then, if you were driving a car and I was blocking your windshield with like a piece of wood or whatever. You're basically allowed to look on both side. So you see what's happening right now. You're allowed to look in the past. You have your rear mirror. But you don't, you're not allowed to look in the front of the car toward what's coming at you. So you'll be going through all kinds of surprise, and you're going to react to the surprise, but you're already going through that surprise. You don't anticipate.
So if you were driving like that, it would definitely cost you more energy because you were not able to anticipate, should accelerate or slow down or just let it go. And it will be very, very uncomfortable for whoever passenger is with you in that car. And I will not mention the damage that you're probably going to create on your car.
When you think about it, all of our HVAC systems are acting like that. Because we don't know the future. So we're, we're following the schedule. We're following actions that are happening already, and we're reacting to these actions by a reaction. And we're trying to stabilize as we go. So knowing the future, suddenly we could start to do preemptive action that would either improve the comfort, or reduce the quantity of kilowatt that our spent to react to an event.
[00:16:47]
Doing the control is the hard part. Because the big issue we have, like with building, when you think about it, is you're not trying to resolve one linear problem. So let's say one linear problem could be, I want to maintain the temperature all the ways within my dead band of the set point. So that would be like a linear problem and you would not need neural network to do that. A classical machine learning tool would do it. But then you want to do, okay, I want to do that, but at the same time, I would like to do it with the minimum quantity of kilowatt hour that I'm spending over time. Oh, but then you have a second problem that you're trying to resolve and you're trying to optimize both of these problems at the same time.
But then, Oh, wait a sec. This is not it. Because the utility, they're charging us on the power factor too. So even though I am saving kilowatt hour, if I'm creating a super peak on the power side, I'm not going to win at the end of the month when I'm going to receive that bill. Actually, that could create a worst case, because the power factor of the bill, it could be very intensive in dollars. So we want also to optimize that kilowatt in power, as a third line that you want to optimize at the same time. So, Oh, this is great. So then we're starting to have a more complex problem that you want to resolve. So you want to optimize these three lines in parallel.
But wait a sec. We don't want to create some cycling on the equipment. Right? Because we will be slowly destroying too fast like a pump or a ventilator. So, so I want to make sure I'm not cycling anything here. I'm treating his equipment with respect and I'm not creating equipment problem down the road. So, Oh, that's a fourth line that I need to optimize in parallel to the first three.
And then your problem is becoming so complex. It would be resolvable. I mean, we would put like probably, you know, you, me and a couple of other engineer, a couple of stats model around a table. We're going to crunch this, maybe with MATLAB, and we would come with the solution. Probably going to take us a few hours. But we would come with that solution of what is the optimal control for the next hour. By the time we do it, of course, that hour is long gone. So whatever we came up as a solution is too late, because we should have done that in half a second if you want to be efficient. And just imagine the cost of all of us around that table. And we would have, of course, to work 24/7, day and night because you know, we need to operate that building and that's only one building.
So it's, it's all feasible by hand, but the power of like a deep, reinforcement learning. And that's the same, the same thinking than the video game. So when the AI is playing against us on a video game, it is quickly understanding what are the rewards, so how to win the game. So if I want to win the game, I need to maximize basically my points and to get points, there are several ways you could get points, right? So quickly the AI is understanding how to get these reward, and how to score maximum value and win the game.
So it's the same, same technique we're using, but we're kind of presenting to the AI as a game, saying we have all of these paths and you need to optimize all of them in parallel. And if you manage to do like a global balance, you're going to win the game. And that's what we call the deep reinforcement learning. And this is what we're applying. And it's producing a better control strategy than the typical control sequence, which was already there in the building. Which is, as you described earlier, you know, based on fixed schedule, which are not always tweaked the right way, which is based on reaction. So it's reacting. Remember, the AI knows what's coming, so he's got a clear advantage on that control sequence because he knows what's going to be the demand for let's say for the chiller in two to three hours. So it could pre-produce water without creating a super peak.
[00:26:03]
So one of the big challenges we had to resolve-, and it was nothing to do with artificial intelligence, was the fact that there is a lot of different controllers out there on the HVAC side. We actually did a count and we went all the way up to about 700 HVAC controlled protocol in the world. Some of them are, you know, were created by company that don't exist anymore. They're not being supported anymore. Some of them were bought by a bigger company and they're trying to support the line and at the same time convince a customer to, to switch over. There's also a lot of different versions, so, so on each of these control protocol there is different version, and a building may necessarily upgrade their control systems. They try to, you know, go as long as they can without upgrading because they want to avoid the expense.
So it's a pure nightmare. You want to interface with that kind of ecosystem, and some of these protocols are completely closed. I mean, there's absolutely no way you could understand that language. There's no documentation. The company does not exist anymore. Good luck to find somebody that knows about it. So we had to create an edge device, a physical edge device. That we could install in the building, and would be able to connect to that existing control system and talk the same language.
So of course, you know, BACnet was the easy one. But think about LonWorks. Think about Modbus. Honeywell is an interesting protocol just to name one. So it's, you need to be able to connect that box and that box is able to talk that language to first of all, discover what are the points, so an auto discovery. Then read the point. And when we read the point for the first time, let me just tell you that it's a, it's a very rare exception when we hit the building, which is already Haystack tag. Most of the building, it's a very weird nomenclature, which is sometime created by the technician that did that set up. We see building, which is, let's say, half in English, half in Spanish. And we see building where people are giving names to the equipment. I mean so instead of calling a fan a fan, they give it the name of a person. You know, this is Andy and it's a fan, but they don't say it's a fan, just this is Andy. So we have to figure out that Andy is a fan. So, you know, nomenclature is a very creative world in the building HVAC control. And so that's a step that, that we need to do manually.
And actually we're working with the NREL in Colorado, and we are creating what we call the Autobot. So the Autobot is a piece of AI which is doing that mapping, that conversion of whatever name they were using in that building to a standard, Haystack tag nomenclature, because it's only when you organize your data in that fashion that the artificial intelligence could understand what it is.
And that is the step that we call mapping, so still manual process in the motion to becoming AI too in terms of the mapping. So if the AI could do about 80% of that mapping, and we still have human doing 20% of the mapping, I think that would be very, very happy if we reached that level of applying the AI for the mapping, naming conversion.
So once that step is done, this edge device connected to the existing controller of the building will start to read the data and send it to the cloud. So we have different way to do that, in term of secure connection. And that reading goes in the cloud where it accumulate in a database, which is specific for that building.
And then we have to wait. We basically have to wait that period of time to accumulate enough data. We're also in the cloud mixing that data coming from the building with detailed, detailed whether data that I mentioned previously. So you want that, that special weather data, which is giving you the wind, the wind direction, the wind gust, the cloud thickness. So we're not just talking about humidity and temperature here. We're really talking about detailed weather, because there's a lot of correlation between what's happening inside the building with sometime driver, like the wind direction and/or cloud thickness, because that's giving you directly the solar radiation intensity, that cloud thickness.
So that weather data is accumulated in parallel on the same timeline than the data points were taken for the building. During that period of time that I mentioned, you know, five, six, seven, eight weeks, in the same season. And it's only then that we have enough data that you could start to apply this prediction.
And for us, this prediction step is very important because it's giving us the quality control that, yes, we could start to do automatic control because the prediction is good. And it's also something you want to keep doing in terms of prediction analysis. Because when you see a degration of the prediction, it's your signal that you should also retrain your AI.
So that's happening in different situations. It could be happening when there's a season change. So you were, you were training yourself, you were operating during the wintertime, and then you were getting in the spring, you're going to start to see new behavior happening in terms of the weather, of course. And that will have a different impact. Switching from heating to cooling is an interesting aspect, and it's happening in the reverse order in the fall, that requires a retraining of your neural network, so they discover new behavior that's happening all along during the first year.
But it also could be a user behavior change. So a tenant is leaving the 10th floor. He's going to another the building. So suddenly the 10th floor is empty, and then then they will do construction on the 10th floor for the new tenant that's moving in. And then there's going to be a new tenant. Right now we are seeing in the COVID-19 crisis, all kinds of behavior change on a tenant side. So there's a tower, which on 30 floors, there's only two floor, which are still occupied, and it's occupied by a government department, which used to work nine to five, now they're in crisis management, so they're working from six in the morning to midnight. So and all of the other floors are now empty. So who's going there to change the entire control sequence to adapt to this new reality of that tower, right? Feasible. Is it being done? Well, the AI automatically recognized there's a big shift happening and that night will retrain itself to this new behavior that it's understanding. And after a few days, we'll be now completely understanding the new setup of the tower and we'll be trying to optimize that new set up.
Well, that retraining is happening in the cloud. So most of the AI is happening in the cloud. And then we have a lot of what we call control algorithm. So we have about 20 something of these algorithm , and they all focus on one part of the system. Either they focus on the entire tower, like we have an algorithm which is optimizing the schedule. So what time should we start the system this morning? What time can we stop the system in the evening if there's a schedule, if it's that type of building. We have other algorithms which are trying to optimize the temperature of the water, of the hot water being supplied into the system at any given time. So same thing on the chiller side. We have other algorithm which are focusing on power peaks. So how can we manage to not create another peak of power that month in the tower.
We have algorithm which are focusing only on air handling unit. So when is the optimal time to start the cooling stage one or cooling stage two? We have other algorithm working only on fan speed if you have variable speed drive. So what is right now my optimal static pressure in that duct at any given time.
And we have to make all of these algorithms work together. So, we need a coach, which is making sure that the team is doing a good teamwork and not working one against the other. So that's the complexity of it. So most of these tactical algorithms are located at the edge device, because the ones that are doing these decision on real time are located at the edge device, but they work in tandem with the crunching happening on the cloud side.
James: So another type of challenge that I thought of as I think about my average building operator, so there's two challenges and they're both pride related. So, when I think about a building operator, I think about one that either has pride in his control sequences or one that has pride in the way that he operates the building without his control sequences, right? So, in the former, it's like, we have, the state of the art control sequences. Our mechanical engineer did a great job. I don't really understand them. I just know they're working really well. And the latter, it's I don't trust control sequences. I operate my building. Right? And how do you guys approach those two separate pride challenges?
Jean Simone: [00:48:37] Yeah, I mean, you're touching probably the most difficult part of any project, right. How do you make the AI work with humans? So, I don't know if you had the pleasure to sit in an autonomous car level four, but, I was told that, I was really not a good person to do that test, because I'm fighting with the AI. I want to control the wheels. I don't trust it, and it's because I've been conditioned with, you know, a lot of years of driving myself. And for me, you put me behind a wheel and I will drive, right? I'm not gonna just cross my arm and look at the show, right?
So, I think anybody, which has, you know, been maintaining a building for, for years and years, and it's becoming your building and you're very proud of what you do, and you should be, it's not an easy job. You know, I think of all these tenants which keeps complaining and you know, issues, and you're trying to maintain that level of happiness, with all of the users of the building, it's an art. So with the tool that you were given, you pushed them and you configured them through a level of which you're quite satisfied. And you did put so many hours to make it, tweak it so it does the best it could with whatever you have. So, and then suddenly you have that piece of AI coming in and saying, Oh no, we're going to go get like a lot more value out of this building. And it could be very frustrating, to see that, wait a sec. You know, do you know how many hours and years that I put into this?
So, and there's an interesting story about this. I was talking to one of these person that we're describing here, which spent his entire career in one tower. And he's about five years away from retirement, and he's the ultimate expert of that building. And he was looking at me, he said, there was nothing the AI could do better than me. And I was like, okay, yeah, interesting discussion. And I said, well, are you going home at night? And he said, yeah, yeah. Yep. Seven to three, sometimes 3:30. Okay. Perfect. So who's-, if you're not here, what happens? Because there's a shopping mall at that bottom of that tower, during the evening, you know, Thursday, Friday night, that's open until nine. You know, what happens in the shopping mall?
And he said, well, if there's a problem, they're going to call me. Okay. And you like that, to be called? No, no. No. Okay. And what happens Saturday? Shopping mall open, right? Yeah. Yep. They will call me if there's a problem. And what happens if there's something happening during the night and then? Oh, I mean, that's not good. So what about the AI is becoming your second in command, and it's managing the ship 24/7, so you could have a better night to sleep, so you, you could go on vacation two weeks somewhere in Europe. You could go for lunch, not being stressed that your pager is going to ring. And he looked at me and said, okay, now I understand.
And he looked at me and said, and maybe if I show the AI a lot of things in the next five years, maybe I could retire with peace of mind. And I said, yeah.
James Dice: [00:51:26] Yeah. And the building owner would have peace of mind as well.
Jean Simone: [00:51:29] Yeah, so, the AI is not here to replace people. AI Is here to give us the ability to do more on our daily basis. It's increasing our capability or capacity to do a lot more and to maybe focus our creativity on the most important things and letting all of the basic stuff being done by the AI, freeing you up some time to do more important things with your expertise.
[00:56:56]
It's interesting because we were talking about that a few weeks back, trying to understand what will be this new world. Last year we were selling more on we will improve the comfort, so, you know, we will reduce issues you have on the comfort side. And a lot of customers were like, it was their key priority. Yes, I want to save money, but I have this tenant on the sixth floor, which keep complaining and we can't stabilize the temperature. And if your AI could fix that, it's worth the money. And there was also a lot of, let's save the planet. If we reduce the energy intensity, I'm going to be able to say that my building consumed less per square. I'm helping the planet. That was really the spin on the sales side.
What we're seeing now is that kind of completely shift now. Now people are, okay, I need to save money. I lost two tenants. They're not paying their bill anymore. My revenue are coming down, so now I need to do whatever I can on the expense line to reduce that one, because the bottom line is suffering like hell.
So we're now becoming like a cost saving tool, and it's probably the only key driver that we need to push forward is you have to perceive us as a no install-, a very low install costs, and I will be able to have a significant impact on that expense line immediately. Well, within the first two months, right? We have to learn about it. But quickly it's going to contribute to that bottom line number, which is now the game, right?
James Dice: [00:58:20] Yeah. Yeah. I mean, I think about how like, you know, it's May 7th. So, you know, at mid March, it was all about how there's no one in the building, right? So can we shut it down completely? Now, as we start to talk about reoccupying, there's people like me that are probably going to be working from home for three more months, I would think. There's not a whole lot of reason for me to go back to the office, but there are quite a few people that are now starting to trickle back in, right? So a solution like this can help keep the energy-, you know, it's not fully shut down, but it can keep things tailored to how many people are actually there.
Jean Simone: [00:58:54] Yeah, I mean, absolutely. I mean, it's-, because of the retraining and the learning, the self-learning, it's adapting quickly to the new configuration. And, and you're right, it's going to be a moving target, right? Some people will go back, not everybody. Take a typical tower and downtown core, probably now it's either completely shut, it's basically on set back 24/7. Or there may be one or two floor because essential services are working on these floor. And then slowly you're going to have this company on another floor starting, but starting with social distancing. So probably half the employees only are going to come back and not the other half, or they're going to do rotation shift or whatever.
So, that control sequence will need to be modulated as a moving target from day to day as this confinement modulation is happening. And suddenly we could have a new research and then the confinement is back quite rapidly. So you need to rechange everything. So we're pretty convinced that the AI will bring a lot of value there. We will continuously trying to reset and basically follow that, that behavior shift, which is happening on a weekly basis. So, you know, once again, it would be feasible to do by sending a control technician to readjust the control sequence on a weekly basis. But I'm not sure how many people will really do that.
[01:00:46]
What I call the BrainBox 2.0, which is, okay, imagine if all of these AI engine, which are operating in their own building, they're not aware there's other AI engine working in a building across the street. Right? So imagine if we were connecting them together and they were kind of becoming aware that there is other AI engine deployed another building in the same downtown core. And what happened if they share information, then they start to work as a team to the benefit of the grid. So, you know, if the grid would be like sending signal and these AI engine , knowing what's happening in the next hours in term of prediction could be able to play a strategy to the benefit of the grid, especially if there's a reward allocated to them, by the utility. So we call this the swarm AI, and it's something that we will start working on pretty soon.
What did you think about these highlights? Let us know in the comments.
Note: transcript was created using an imperfect machine learning tool and lightly edited by a human (so you can get the gist). Please forgive errors!
James Dice: [00:00:00] Hello friends. Welcome to the Nexus, a smart buildings technology podcast for smart humans. I'm your host, James Dice. If we haven't met before, I write a weekly newsletter on the same topic. It's also called Nexus. Each week I share what I've learned, my opinions, and what I'm excited about in the quickly evolving world of intelligent buildings. Readers have called Nexus the best way to stay up to date on the future of this industry without all the marketing fluff. You can check it out and subscribe at nexus.substack.com or click the link in the show notes.
Since starting the Nexus newsletter, many of you have reached out to me wanting to talk shop, and we have. After a few weeks of those wonderful conversations, I realized I needed to record and share them with our growing community. So here we are. The Nexus podcast is born. This is our chance to explore and learn with the brightest in our industry together.
One more quick note before we get to this week's episode. I'm a researcher at the National Renewable Energy Laboratory, otherwise known as NREL. All opinions expressed on this podcast belong solely to me or the guest. No resources from NREL are used to support Nexus, and NREL does not endorse or support any aspect of Nexus.
Okay. Here we go. Episode 8 is a conversation with Jean-Simon Venne, the CTO and Co-Founder of BrainBox AI, an advanced supervisory control platform for autonomous buildings. We do a deep dive on BrainBox AI, including how they use artificial intelligence to save energy and to adapt building operations to occupancy patterns of the pandemic.
We talk about the problem with existing control systems and how AI helps supplement them. We talk about how AI and humans can be a great team and how the BrainBox team is approaching that teamwork. Finally, we talk about how autonomous buildings can lead to an autonomous and cleaner electric grid. And much, much more.
This episode of the podcast is directly funded by listeners like you who have joined the Nexus Pro membership community. You can find info on how to join and support the podcast at nexus.substack.com. You also find the show notes, which has links to BrainBox AI's website and Jean-Simon's LinkedIn page. Without further ado, please enjoy Nexus Podcast Episode 8.
Jean-Simon, welcome to the show.
Jean Simone: [00:02:21] Thank you. Good morning.
James Dice: [00:02:23] Yeah. So why don't you give us your personal background, and then a little bit about BrainBox.
Jean Simone: [00:02:28] Yeah, absolutely. So, I've been working in technology pretty much all my career, evolving with the technology. As you remember in the eighties, it was not the way it is today. So I kind of followed that path. I would say that my previous said adventure before BrainBox was all about optimizing energy in buildings, but I would say the old way. No offense here, but it was about, you know, big retrofit of existing equipment, let's put a new boiler, more efficient, let's try to put some economizer on existing equipment. So always big construction project, a lot of capex, time to execute a project, you know, engineering plan and specification, manage the construction, commissioning. You're easily talking 12, 18 months. And then you hope, like everything is tweaked the way it should and the saving will be, will be at the meeting point. Right. And, and usually they are, and, you are, you're all happy and you leave, you leave that project thinking job done. And that adventure was actually more on the ESCO side, so we were paid a percentage of the saving over the years.
Usually first year, all good. You know, and then the drifts start this, this infamous drift that you see happening and you can't, you can't really put your finger on exactly why it's happening, but it's happening. And it's a very slow drift. And of course, the first year, it's not really an issue because it's, it's such a small percentage that you barely notice it. But at the year two, when you have that measure and verification annual meeting, Suddenly on the year two, it's becoming more like annoying. You see that saving slowly disappearing like the snow in the spring, and you, you start to wonder, you know, gee, this, this ESCO project is going on for another four or five years. So you, you anticipate, as you know, as an engineer, you start to anticipate what this meeting will be in two, three years. And of course it's following that path, so by the time you were on year four, it's a very unpleasant meeting ,where it's always about debate and why the savings are not as good as they were the first year, et cetera. So it generates a lot of frustration.
And when you, you dive into the nature of the problem, you say, oh wow. I mean, this is thousands of small action that on a day to day operator are taking, and it's slowly de-focusing the system and it's kind of making-, trying to fix some problem, you make other problem showing up. And we should do continuous commissioning. Oh, let's do that. Actually, I went there, so let's do continuous commissioning. It's an after sales service type of thing, we're going to charge that much per month. We're going to build this control center with top gun control people, mechanical engineer. They will remotely make sure the system is always perfectly set up, and we're going to offer that service.
And then you hit that wall of, well, first of all, customer building operator or building owner, not really inclined to pay for such a service, don't understand why should I pay for that? I mean, why is it degrading? So it's, it's a very hard sell, but then you have another systemic problem, which is there's just not enough control expert or control engineer, mechanical engineer on the market available to offer that service on a massive scale. It's just not possible. Even if the customer were willing to pay for it, you would not find all of the people to offer that service for the entire building stack. So, so you're check mate before you start. And that's where the I'd start to wonder, wait a sec. I mean, can algorithm do that continuous commissioning instead of people?
And then you kind of fixed, with one stone, you fixed that double barrel problem, which is an algorithm would be costing a lot less than putting human 24/7 in a control center. They don't take breaks. They don't go on vacation. And at the same time, so the service will be very, very low cost and the same time you could scale, which was literally impossible with humans.
So that, that's how it happened in term of, that put me on that path to, maybe we should use technology to fix that problem.
James Dice: [00:06:46] Cool. Yeah. You just described like most of my career there. Yeah, I've done a lot of that old way style of energy efficiency. Yeah. From
Jean Simone: [00:06:55] Frustrating. It's always very frustrating. Yeah.
James Dice: [00:06:58] Yeah, definitely. And yeah, my most recent role was to develop a monitoring based commissioning technology stack, and then also the services to wrap around it, so I'm very familiar with everything that you just described.
So, cool. So how did, how did you then take that experience into starting BrainBox? And when was that?
Jean Simone: [00:07:19] It was in the 2016, looking at the autonomous car. I'm kind of fascinated by it. So you know, they, they kind of installed this, this, equipment on the car, which is, we look at all that equipment, it's actually costing more than the car itself. And they do that to gather or to generate the data they need, but once it's done and they have all of that data about the surrounding of the car in real time, then they basically use that data to predict that future of where all of these moving objects around the car will be in five seconds and 20 seconds and 30 seconds. And knowing with a very high accuracy what this immediate future is.
Then what is the term of operational research? What is the optimal trajectory or the safest trajectory into that future that I should follow? So then, then that become your control strategy for the car itself. Like should you go left, right, or a right, accelerate, slow down. And they're then doing that with incredible success.
I mean, it's a very complex environment. They have milliseconds to calculate and call the shots, and they're doing it. Actually there was one stats that completely flabbergasted me, it was that they were able to learn to behaviors of squirrels and cats. And if you think about it, then we all notice it, like the squirrels have this behavior as they, they're not sure if they're going to go, they don't go, they go, and then suddenly they go. And then when they realized that there's something that is not going the way they want, they stop in the middle of the street. And they question themselves again, which is the worst possible thing, right? And and as a human, we learn over time that the squirrel will probably do that. He's not going to continue his run. He's probably going to stop trying to assess a situation, and that's how I'm going to hit it. So we anticipate that. So they managed to do the same on the AI side, anticipating the squirrel behavior.
And I look at that and like, wow. If they could do this, we should be able to do the same in HVAC. We have a lot less moving part, a lot less risk. And on top of it, I have the luxury of time. I don't have millisecond to calculate and call a shot. I have minutes. Because, you know, changing the temperature in a room, takes what, 10, 20 minutes, depending on the size of that room? So I got this inertia on the thermodynamic side, which is giving me the luxury of time to calculate. So there's no reason it should not work. So that was the ticking of the idea of saying, oh, we should do that. Use the existing AI techniques, use an autonomous car, and use it for doing an autonomous HVAC. It's how it's really started around 2016.
James Dice: [00:10:05] Cool, okay. I want to dive into the autonomy piece a little bit. Let's back up though, before we get there. Why do we need autonomous buildings? Like what are the other problems with controls specifically, you know, besides the fact that buildings aren't energy efficient and they drift, like you talked about earlier, what are the issues with the actual old style of controls?
Jean Simone: [00:10:29] So, yeah, so there's the entire continuous commissioning aspect, which we already spoke about. And yes, at the beginning, that was the target. But then we realized that if we're able to predict the future of what will be happening in that building with a super high accuracy or about 99%, we basically know the future. And when we realized that the deep learning algo were, were giving us that accuracy for hours and hours ahead, so we knew that in three hours, that room will be too hot. It will hit the cooling set point in three hours, in four hours, in six hours, and we realized that we had that, that luxury to basically know the future of that building with a very, very high accuracy. You went like, wow. I mean, we could do more than continuous commissioning.
We could basically design our control strategy right now to have a better future. And it's, it's kind of going into this, this concept of, when you think about it, all of our HVAC right now, they're reactive. And even if you use a PID loop, you're reactive. So there's something which is happening now, and you're reacting to it. But you're already into that climb or in that descent, you're already going through that event. You're not into the past or looking at the future. You're really into the present. And all of our HVAC systems are reactive. I mean, they are programed which if this value trigger this action until you have this other value, then you could release or you could stop whatever action you were doing.
The entire logic of our programmation is like that, is reactive. So it's kind of similar, then, if you were driving a car and I was blocking your windshield with like a piece of wood or whatever. You're basically allowed to look on both side. So you see what's happening right now. You're allowed to look in the past. You have your rear mirror. But you don't, you're not allowed to look in the front of the car toward what's coming at you. So you'll be going through all kinds of surprise, and you're going to react to the surprise, but you're already going through that surprise. You don't anticipate.
So if you were driving like that, it would definitely cost you more energy because you were not able to anticipate, should accelerate or slow down or just let it go. And it will be very, very uncomfortable for whoever passenger is with you in that car. And I will not mention the damage that you're probably going to create on your car.
James Dice: [00:12:56] Yeah, that sounds terrifying.
Jean Simone: [00:12:58] But that, when you think about it, all of our HVAC systems are acting like that. Because we don't know the future. So we're, we're following the schedule. We're following actions that are happening already, and we're reacting to these actions by a reaction. And we're trying to stabilize as we go. So knowing the future, suddenly we could start to do preemptive action that would either improve the comfort, or reduce the quantity of kilowatt that our spent to react to an event.
And when you're at full throttle, let's say it's super hot in the summer or it's super cold in the winter, and the system are at maximum capacity. Knowing the future is not really helping you. You're already at maximum capacity. You will be running at maximum capacity for the next three hours. You're not going to save anything, but there's a lot of other periods during the year where you're not running at maximum capacity of equipment and knowing what's going to happen, and you could do preemptive action, which will save some money, and at the same time will reduce the discomfort that these actions and reactions are creating.
James Dice: [00:13:55] Yeah, like 99% of the time. Other times besides peak times.
Jean Simone: [00:14:00] Yeah. Yeah. Whatever is the building and how it was designed, right?
James Dice: [00:14:03] Yeah. Yeah, so I'm writing a piece right now for Nexus about how the old style of supervisory controls are really just quite dumb when you think about it. You have a system that requires a very expensive controller, right? And then you have a server that groups all of those very expensive controllers together. So you're talking about many, like tens of thousands of dollars at this point, and now you have this system that basically is a database that doesn't work that well, a trending device or visualization device for those trends that really doesn't really help you that much. You have some alarms that no one pays attention to. You have some schedules that are easily over-ridden and really aren't that sophisticated also. And you have some graphics that are usually wrong, right?
So it's just like this system that is really 10, 20, I don't know, maybe you could say 30-, it's like a mainframe system from before I was born. Right. So, uh, yeah. So-
Jean Simone: [00:15:04] Careful, I played on them.
James Dice: [00:15:07] Yeah. So I mean, yeah. So compare that, like I've been saying, compare that to the iPhone 11 I had just got a couple of weeks ago, and it's just night and day. You have a very dumb system versus today's smart technology.
Jean Simone: [00:15:23] Yeah. And I mean the beauty is interesting because the people already paid a lot of capex to install these controlled systems in the building, which are generating an incredible quantity of data. But that data is used on the incident of the moment, maybe kept for a few weeks in logs, but then is overwritten, right. And there is a very, very few buildings which are keeping like years of history for the granularity of all of the data point, with the very small reading interval. But the data, I mean, you already paid to generate that data, so that the big heavy lifting on the capex is done.
Think of the autonomous car. I mean equip a car to be autonomous. The big heavy lifting is all of the electronics that you need to onboard on that car. And that's the barrier to have a lot more autonomous car. I mean, who has the money to buy these type of car? The electronic onboard is more expensive than the car itself.
So, but on the building site, it's already done. People already paid huge amount of money to build these control system, which are generating a fascinating amount of data. But we're not using fully the potential of that data to create a lot more value with it.
James Dice: [00:16:34] Totally. Alright. So I want to dive into the product now. So first thing, kind of paint besides prediction, take us through kind of like a day in the life of an autonomous building.
Jean Simone: [00:16:47] So there's really the two big component once it's in operation, there's, there's the predictive. So you want to use these deep learning model neural network. Think of LSTM as an example, which gives you this very high accuracy in term of prediction, once you have enough data so that-, you know, one of the issues is you need to accumulate enough historic data to start running these systems. So easily, you know, five, six, seven weeks of data within the same season. So it's not something you could kick in within two days and say, well, ah, we have AI in this building. So you need to connect, you need to accumulate that data, and only when you have a critical mass within the same season, then you could really kick in, deploy these type of tools. So there is barrier, and you have to respect them.
So this prediction is giving you this vision of the immediate future. So on the air side, you really want to tackle the next three hours. Even though we have more, the neural network will give you a high accuracy of what's going to be happening on the air side for the next, six, seven, eight, nine, and even few days. This is how powerful it is. But on the air side control, you don't really need more than three hours ahead of you.
On the water side, think of, you know, boiler chiller water networks, then you really want more six to eight hours, because the inertia of producing water at a certain temperature is something which is prepared in a longer timeframe than, than the air side, makeup air if you want. So that prediction is the first fundamental part.
So doing prediction, on the AI, we have a saying which says, doing the prediction is the easy part. Doing the control is the hard part. Because the big issue we have, like with building, when you think about it, is you're not trying to resolve one linear problem. So let's say one linear problem could be, I want to maintain the temperature all the ways within my dead band of the set point. So that would be like a linear problem and you would not need neural network to do that. A classical machine learning tool would do it. But then you want to do, okay, I want to do that, but at the same time, I would like to do it with the minimum quantity of kilowatt hour that I'm spending over time. Oh, but then you have a second problem that you're trying to resolve and you're trying to optimize both of these problems at the same time.
But then, Oh, wait a sec. This is not it. Because the utility, they're charging us on the power factor too. So even though I am saving kilowatt hour, if I'm creating a super peak on the power side, I'm not going to win at the end of the month when I'm going to receive that bill. Actually, that could create a worst case, because the power factor of the bill, it could be very intensive in dollars. So we want also to optimize that kilowatt in power, as a third line that you want to optimize at the same time. So, Oh, this is great. So then we're starting to have a more complex problem that you want to resolve. So you want to optimize these three lines in parallel.
But wait a sec. We don't want to create some cycling on the equipment. Right? Because we will be slowly destroying too fast like a pump or a ventilator. So, so I want to make sure I'm not cycling anything here. I'm treating his equipment with respect and I'm not creating equipment problem down the road. So, Oh, that's a fourth line that I need to optimize in parallel to the first three.
And then your problem is becoming so complex. It would be resolvable. I mean, we would put like probably, you know, you, me and a couple of other engineer, a couple of stats model around a table. We're going to crunch this, maybe with MATLAB, and we would come with the solution. Probably going to take us a few hours. But we would come with that solution of what is the optimal control for the next hour. By the time we do it, of course, that hour is long gone. So whatever we came up as a solution is too late, because we should have done that in half a second if you want to be efficient. And just imagine the cost of all of us around that table. And we would have, of course, to work 24/7, day and night because you know, we need to operate that building and that's only one building.
So it's, it's all feasible by hand, but the power of like a deep, reinforcement learning. And that's the same, the same thinking than the video game. So when the AI is playing against us on a video game, it is quickly understanding what are the rewards, so how to win the game. So if I want to win the game, I need to maximize basically my points and to get points, there are several ways you could get points, right? So quickly the AI is understanding how to get these reward, and how to score maximum value and win the game.
So it's the same, same technique we're using, but we're kind of presenting to the AI as a game, saying we have all of these paths and you need to optimize all of them in parallel. And if you manage to do like a global balance, you're going to win the game. And that's what we call the deep reinforcement learning. And this is what we're applying. And it's producing a better control strategy than the typical control sequence, which was already there in the building. Which is, as you described earlier, you know, based on fixed schedule, which are not always tweaked the right way, which is based on reaction. So it's reacting. Remember, the AI knows what's coming, so he's got a clear advantage on that control sequence because he knows what's going to be the demand for let's say for the chiller in two to three hours. So it could pre-produce water without creating a super peak.
There was a good example that I like to bring forward in a very high rise, a very big high rise in Montreal. One of the issues they had is that their chiller is working at full, full throttle. A Beautiful day outside in the middle of July, very hot, and it's sunny, and then suddenly these clouds move in around 2:00 PM, very thick cloud, and suddenly it become very dark.
So within 15 minutes, there's hundreds of tons of chilled water that are over capacity in that building. So they start to shut down the systems and lower the production because they're in full excess capacity. But the AI knows these clouds are coming, because right now, when you're taking this pilot weather data, and you know exactly the movement of the cloud and the cells, and the thickness of the cloud, and you know where are they going to be. And so we could anticipate that you're going to have like 200 tons of chilled water in excess by 2:00 PM this afternoon. And so you could start to glide it in advance, and slowly lower your production on the cold water in anticipation of that, and not reacting to an event which is happening suddenly.
So that's just an example, but it gives you a very smooth curve, knowing the future and it's saving kilowatt here and there, all day long.
James Dice: [00:23:43] Totally. That's a great example. So I want to circle back real quick, so comparing autonomous buildings to autonomous cars, it seems like it's a different output problem. So with a car, it's like I am either going to brake, I'm going to accelerate, I'm gonna turn left, I'm gonna turn right. There's, there's just limited output, right? Versus a building, it seems like , like you said, we're optimizing for more variables. Is that the right way to think about it?
Jean Simone: [00:24:08] Yeah, I would say that you're right, there's more variables on the HVAC than on the autonomous car, but we have less moving part. So, so you're right, I mean the autonomous car has less variables that they have to treat, but they have so many moving objects around them that it's making the problem very complex.
On our side, we have very few moving object, it's actually a very static environment, but we have a lot more variable to consider to figure out what is the thermodynamic equation that will be happening here. And then again, I mean, the AI does not, artificial intelligence does not know the thermodynamic handbook that we all study in engineering, because it's only data driven. But it's figuring out the: if this happened, this will happen. It's figuring all of these action reaction that you see typically in the building, and it's becoming super customized on one building. So once it's been learning on one building, it's what we call completely burn.
That model cannot be taken and used in another building because each building is so different on the way it's behaving in term of the thermodynamic side that, that really, once it's learned that building it, yeah, it's only good for that building.
James Dice: [00:25:18] Okay. And you also mentioned this acronym LSTM. What's that stand for?
Jean Simone: [00:25:23] Yeah, that's one of the-, there's a lot of different neural networks, that are out there that could be used as one of the models. We're using several types. One of the things we do is, is we basically when we do the learning part is we test different neural network model and we figure out which one is the best fit for this specific building.
James Dice: [00:25:41] Hmm. Got it. Okay, cool. So tell us about kind of, the rest of the stack. So I'm assuming all those AI algorithms are in the cloud, right? So how does the, how do you get the product? I'm assuming there's a box, right? So it's called BrainBox. I'm assuming there's a box. So, tell us about the entire stack from the building up to the cloud.
Jean Simone: [00:26:03] Yeah. So one of the big challenges we had to resolve-, and it was nothing to do with artificial intelligence, was the fact that there is a lot of different controllers out there on the HVAC side. We actually did a count and we went all the way up to about 700 HVAC controlled protocol in the world. Some of them are, you know, were created by company that don't exist anymore. They're not being supported anymore. Some of them were bought by a bigger company and they're trying to support the line and at the same time convince a customer to, to switch over. There's also a lot of different versions, so, so on each of these control protocol there is different version, and a building may necessarily upgrade their control systems. They try to, you know, go as long as they can without upgrading because they want to avoid the expense.
So it's a pure nightmare. You want to interface with that kind of ecosystem, and some of these protocols are completely closed. I mean, there's absolutely no way you could understand that language. There's no documentation. The company does not exist anymore. Good luck to find somebody that knows about it. So we had to create an edge device, a physical edge device. That we could install in the building, and would be able to connect to that existing control system and talk the same language.
So of course, you know, BACnet was the easy one. But think about LonWorks. Think about Modbus. Honeywell is an interesting protocol just to name one. So it's, you need to be able to connect that box and that box is able to talk that language to first of all, discover what are the points, so an auto discovery. Then read the point. And when we read the point for the first time, let me just tell you that it's a, it's a very rare exception when we hit the building, which is already Haystack tag. Most of the building, it's a very weird nomenclature, which is sometime created by the technician that did that set up. We see building, which is, let's say, half in English, half in Spanish. And we see building where people are giving names to the equipment. I mean so instead of calling a fan a fan, they give it the name of a person. You know, this is Andy and it's a fan, but they don't say it's a fan, just this is Andy. So we have to figure out that Andy is a fan. So, you know, nomenclature is a very creative world in the building HVAC control. And so that's a step that, that we need to do manually.
And actually we're working with the NREL in Colorado, and we are creating what we call the Autobot. So the Autobot is a piece of AI which is doing that mapping, that conversion of whatever name they were using in that building to a standard, Haystack tag nomenclature, because it's only when you organize your data in that fashion that the artificial intelligence could understand what it is.
And that is the step that we call mapping, so still manual process in the motion to becoming AI too in terms of the mapping. So if the AI could do about 80% of that mapping, and we still have human doing 20% of the mapping, I think that would be very, very happy if we reached that level of applying the AI for the mapping, naming conversion.
So once that step is done, this edge device connected to the existing controller of the building will start to read the data and send it to the cloud. So we have different way to do that, in term of secure connection. And that reading goes in the cloud where it accumulate in a database, which is specific for that building.
And then we have to wait. We basically have to wait that period of time to accumulate enough data. We're also in the cloud mixing that data coming from the building with detailed, detailed whether data that I mentioned previously. So you want that, that special weather data, which is giving you the wind, the wind direction, the wind gust, the cloud thickness. So we're not just talking about humidity and temperature here. We're really talking about detailed weather, because there's a lot of correlation between what's happening inside the building with sometime driver, like the wind direction and/or cloud thickness, because that's giving you directly the solar radiation intensity, that cloud thickness.
So that weather data is accumulated in parallel on the same timeline than the data points were taken for the building. During that period of time that I mentioned, you know, five, six, seven, eight weeks, in the same season. And it's only then that we have enough data that you could start to apply this prediction.
And for us, this prediction step is very important because it's giving us the quality control that, yes, we could start to do automatic control because the prediction is good. And it's also something you want to keep doing in terms of prediction analysis. Because when you see a degration of the prediction, it's your signal that you should also retrain your AI.
So that's happening in different situations. It could be happening when there's a season change. So you were, you were training yourself, you were operating during the wintertime, and then you were getting in the spring, you're going to start to see new behavior happening in terms of the weather, of course. And that will have a different impact. Switching from heating to cooling is an interesting aspect, and it's happening in the reverse order in the fall, that requires a retraining of your neural network, so they discover new behavior that's happening all along during the first year.
But it also could be a user behavior change. So a tenant is leaving the 10th floor. He's going to another the building. So suddenly the 10th floor is empty, and then then they will do construction on the 10th floor for the new tenant that's moving in. And then there's going to be a new tenant. Right now we are seeing in the COVID-19 crisis, all kinds of behavior change on a tenant side. So there's a tower, which on 30 floors, there's only two floor, which are still occupied, and it's occupied by a government department, which used to work nine to five, now they're in crisis management, so they're working from six in the morning to midnight. So and all of the other floors are now empty. So who's going there to change the entire control sequence to adapt to this new reality of that tower, right? Feasible. Is it being done? Well, the AI automatically recognized there's a big shift happening and that night will retrain itself to this new behavior that it's understanding. And after a few days, we'll be now completely understanding the new setup of the tower and we'll be trying to optimize that new set up.
Well, that retraining is happening in the cloud. So most of the AI is happening in the cloud. And then we have a lot of what we call control algorithm. So we have about 20 something of these algorithm , and they all focus on one part of the system. Either they focus on the entire tower, like we have an algorithm which is optimizing the schedule. So what time should we start the system this morning? What time can we stop the system in the evening if there's a schedule, if it's that type of building. We have other algorithms which are trying to optimize the temperature of the water, of the hot water being supplied into the system at any given time. So same thing on the chiller side. We have other algorithm which are focusing on power peaks. So how can we manage to not create another peak of power that month in the tower.
We have algorithm which are focusing only on air handling unit. So when is the optimal time to start the cooling stage one or cooling stage two? We have other algorithm working only on fan speed if you have variable speed drive. So what is right now my optimal static pressure in that duct at any given time.
And we have to make all of these algorithms work together. So, we need a coach, which is making sure that the team is doing a good teamwork and not working one against the other. So that's the complexity of it. So most of these tactical algorithms are located at the edge device, because the ones that are doing these decision on real time are located at the edge device, but they work in tandem with the crunching happening on the cloud side.
James Dice: [00:33:58] Yeah. So all the heavy lifting is happening in the cloud.
Jean Simone: [00:34:01] Yeah, because of the processing power that we need.
James Dice: [00:34:04] Yeah. I think what you just described was one the answers to one of my follow up questions, which was, if you lose internet, that local edge device has basically the latest algorithms and/or control sequences in our industry's terminology. It has the latest and greatest from the cloud. And it's getting those every couple minutes as an update, or how does that work?
Jean Simone: [00:34:26] Yeah, we try to standardize. On the air side we are, we're standardizing around five minutes. So we're reading the point every five minutes, we're calculating that prediction every five minutes, and then we're deciding what is the optimal control strategy every five minutes. And then we're waiting another five minutes through to reassess the situation.
We decide to work at the zone level. So our fundamental calculation block is the zone level. And when we want to know what is my strategy for the air handling unit, which let's say might be serving six zones, we basically aggregate the prediction and a control strategy for all six zones together. And then you have your control strategy and your prediction for that air handling unit. And we keep backing it up. If you want to know what's going to be your chiller strategy, well you back up from that valve, which is opening the cooling for that air handing unit, you're backing it up to aggregate all of your air handling units in the building to know basically, what is your best strategy for your chiller at any given time.
So we figure that structuring it like this is giving us the ability that really we don't care the size of the building. So it could be a retail with two zones, or you know, two rooftop, two zone, a small retail store. We're managing them, the two, and it's pretty simple. But a big high rise would be just like maybe 300, 400 zone, aggregateed together like Lego block. So on our side it's just more zone and it's not more complicated. It just takes more time to onboard or to configure it. So that's how we kind of build it.
On the water side, that five minute cycle is too long. So you want to have more like a minute cycle, because that latency of five minutes could become a problem on the water side. So you really want to, you want to more have like a minute cycle when you're playing on the liquid side.
James Dice: [00:36:22] Okay. And what does happen if you lose connection with the cloud?
Jean Simone: [00:36:26] Yeah, sorry, I forgot the main, the main question. Right. So, yeah, so as like autonomous car, the safeties are more important than the action , than your control strategy. Because the last thing you want is to have artificial intelligence which is starting to take the wrong decision and basically controlling your building, right? So, so you need to put a lot of safety in place to monitor if the AI is doing-, first of all, is the AI working? Are we extracting the data? Are we getting the data? Do we have that communication with the cloud? Is the quality of the decision of the algorithm the right quality?
So you need to put basically other algorithm that are checking if the quality that the decision being taken is the appropriate quality, because you want to detect, through all of these safety, if you have a problem. And if you have a problem, you want to de-engage the AI. So one of the problem could be, as you mentioned, that we lose the communication with the cloud. So we need to detect that, and we need to start to monitor that. And basically if the communication is not coming back within a few seconds, you want to start to de-engage.
So de-engage means that you're starting to revert all of your action in a very slow fashion. So we want to have a slow, smooth landing, and you want to basically give back the control to the existing control sequence, which is still there in the background. It's still doing its job, and you want to basically put it back in control of the control sequence. So you're basically de-engaging all of our action and reverting back to that control sequence. And once the communication is reestablished, and even then you want to make sure that it's stable to avoid cycling on the communication problem, then you want to start to reengage automatically.
So this safety, you will find that in an autonomous car, it is actually more safety, about how you're gonna drive that car than there is control strategy of the car, so we have exactly the same situation where we have more safety, especially on the edge device and on the cloud than we have scripts doing the control itself.
James Dice: [00:38:39] Hmm. Yeah, okay. So it's basically when you have full communication you're able to override all of the controllers and send set points and commands to them, basically. And then when you don't have communication, you're then just removing that priority over the set points they already have, basically. Okay, cool.
Jean Simone: [00:38:59] Yeah, yeah.
James Dice: [00:38:59] So a lot of different ways I want to go with that. That's a fascinating view at your stack and how all the things work. So thank you.
So one of the things that strikes me is when I try to-, so shout out to Maddie. Maddie is the one that edits the podcast and helps me produce these. And one of the things that her and I talk about is like, what if I'm explaining this to someone who's new to the industry? And a lot of times it's very difficult. That's why you're laughing so-.
Jean Simone: [00:39:26] Oh, yeah.
James Dice: [00:39:26] So I think what you just described, while it's super high tech, it's, for someone that's just out of school, right, it's not that high tech. Right? It seems like, like the question that comes to mind for someone that's new is like, why weren't buildings doing this already? And one of the things that I like to talk about is all of the enabling technologies that enabled us to, or you guys to build this impressive stack of technologies.
And so I think, I don't want to gloss over the fact that all these are actually pretty new, for buildings at least, right? So you talked about the edge device, which I know you guys use Raspberry Pi's. You talked about BACnet. I mean, that's not that new, but it's also, it is kind of new, right? It's new in terms of the-, the control contractors and OEMs are now being forced by owners to use BACnet or other open protocols. And that's relatively new, right? You guys use cell modems, and so that whole infrastructure you're building on top of that. All of these AI algorithms, so that's not something I'm familiar with, but when did those start coming on the map, in the advent of all technologies in the cloud and that sort of thing?
Jean Simone: [00:40:35] Yeah. Before going on the AI side, I might just back up because you're absolutely right. I mean, just think of storage, data storage. I mean, 10 year ago, 15 year ago, it was so expensive. That if I would have come to any CFO and say, you know what, you have a tower, there's about 40,000 point of HVAC data, I'm going to be collecting a reading every five minutes of that, and I'm going to accumulate it, and then I'm going to use that, you know. The cost of storage, incremental because I keep everything, you know, you want to keep everything when you do AI. We're probably going to have shut down the project right there, because the CFO would have said, are you crazy? It's way too expensive storage of data. But today, cost of storage of data is becoming so low that actually, we're collecting all of the points of the building even though there is probably a few thousand point that we're pretty convinced that we will never need, we're keeping them anyway, because we might discover in two, three years that finally we did find a usage for that data point, and we're very happy to have collected since the beginning. But I mean, the price of storage right now, it's really not an issue.
You mentioned the cellular connection, yes. We're using an IoT SIM card from, from AT&T and Bell Canada. The cost per month is now becoming low enough that, you know what, let's do it. But not long ago, that's cell connection monthly recurring costs plus transmission costs, right, would have probably killed the project right there. Not viable.
So, and then you get into the CPU, I mean the GPU, the capacity to crunch data. I mean, I'm telling my kids that, you know, their cell phone is more powerful than the most powerful computer we had at my university in the eighties. And I showed them a picture of that computer. It had a name, and we were allowed probably 15 minutes on it per week. And you had to justify that it was a worthwhile calculation, that you really need access to it. So when you think about that, I mean. It would not have been possible before to do everything we're doing right now.
So, yeah, I mean, a neural network is a new thing. I mean, a lot of people are talking on the AI about, you know, but it is, it's a new thing. I mean, it's just really flourished in the last six, seven years. Right? Before that, it was not on the radar. I mean, it was being talked about since the eighties. And you, when you think of the, Yoshua Bengio, Lacan, I mean, they were pushing for that for the last 30 years, but nobody was listening to them. Actually they were told that, you know, they were wrong many times. And it's only in the last decade that they had their glory moment where everybody turned back to them and said, wow, you were right. This is the way to go. This is how we're going to do deep learning, deep reinforcement learning, and it's been exploding in the last decade now.
And there's so many possible applications that, I mean, we're just starting to see a real application hitting the market, which are bringing a lot of value, and it's creating new business model, either for startup or for large existing company, but it's just starting to basically shape a new business model, and bringing us into a new world, in term of what can we do better. So, we're really in that stream of applying neural network and trying to create new business model with it.
James Dice: [00:44:11] Fascinating. Yeah. I love that story of like, why now? And it just, all of those enabling technologies coming together, it's pretty awesome.
So, okay, let's talk about, challenges. So, and I've been in a lot of buildings have done a lot of projects, like you talked about earlier, the kind of the old way. And one of the things that, when I think about if I were to go, you know, knock on someone's door and, and sell this technology, a couple of challenges come to mind. So, what about buildings that sort of are in physically poor shape? So, you guys have pulled a lot of data, but what if I can't trust that data? What if sensors need to be calibrated? And what if valves are stuck? What if dampers are stuck? How do you guys approach that, non-software side of things?
Jean Simone: [00:44:57] Yeah. So I mean, and it's the case in every building, right? I mean, it's the reality we're in. So we're, we're limited by the state of the building on the control side. So, if the data is, not calibrated right, you know, we, we see these things, you know, this room is now at 600 fahrenheit. So we have to discard all of these data points, and we're basically blind. So when we're blind in some area of the building, we basically do not engage any type of control in that section of the building. And we, of course, we will notify the building operator, the building manager, the building owner, that, you know, I mean, are you aware that this is a list of points which are either not working anymore or decalibrated. These are equipment which are not working anymore. You mentioned the damper, that's very frequent. You're sending the signal to the activator to open the damper, but the damper is not moving. It's just completely jammed. So I mean, we provide this to the building operator, owner, and suggesting that maybe they should call their control contractor and get them repaired, right? But then it's up to the customer to decide if they want to do it or not. In the meantime, until it's fixed and if it gets fixed, you know, that could take times, we basically removed these area, and we just control the area that we have valid and good data, a reading. So it will create kind of the Swiss cheese, where we're controlling maybe, some of the floor, but not all of the floor, some of the area, but not all the area. And we do as much as we can with the input we get. We can not change, we cannot fix that problem. It's really a building owner, building operator decision at that point.
James Dice: [00:46:39] Yeah. I think it's just like any other thing, any other thing in buildings, is it requires not just the technology, but also the humans and the processes to maintain and maximize or optimize the service itself. So I don't think you guys are going out there and saying, you know, this is a panacea, right? It's, it's not, it's not going to fix everything on its own in its own silo. It's just like everything else that needs to be part of, an ongoing strategy for the building owner.
Jean Simone: [00:47:07] Yeah, we actually try to say, you know, there's no magic here. It's pure maths. So it's garbage in, garbage out. I mean, we're not gonna fix these type of problems.
James Dice: [00:47:16] Totally cool. Yeah, and I think, one of the things that I feel like our audience here, people that are listening to this can get out of this, is that, it's not replacing the other types of building optimization that are out there. Right? So if you think about fault detection, this plays hand in hand with fault detection in that, I mean, maybe these are better control sequence, so there'll be less control sequence related faults, ideally, right? But the physical faults will help-, like when the physical faults are fixed, the control sequences will be able to control more of the building and it'll be able to do better, a better job. So, cool.
So another type of challenge that I thought of as I think about my average building operator, so there's two challenges and they're both pride related. So, when I think about a building operator, I think about one that either has pride in his control sequences or one that has pride in the way that he operates the building without his control sequences, right? So, in the former, it's like, we have, the state of the art control sequences. Our mechanical engineer did a great job. I don't really understand them. I just know they're working really well. And the latter, it's I don't trust control sequences. I operate my building. Right? And how do you guys approach those two separate pride challenges?
Jean Simone: [00:48:37] Yeah, I mean, you're touching probably the most difficult part of any project, right. How do you make the AI work with humans? So, I don't know if you had the pleasure to sit in an autonomous car level four, but, I was told that, I was really not a good person to do that test, because I'm fighting with the AI. I want to control the wheels. I don't trust it, and it's because I've been conditioned with, you know, a lot of years of driving myself. And for me, you put me behind a wheel and I will drive, right? I'm not gonna just cross my arm and look at the show, right?
So, I think anybody, which has, you know, been maintaining a building for, for years and years, and it's becoming your building and you're very proud of what you do, and you should be, it's not an easy job. You know, I think of all these tenants which keeps complaining and you know, issues, and you're trying to maintain that level of happiness, with all of the users of the building, it's an art. So with the tool that you were given, you pushed them and you configured them through a level of which you're quite satisfied. And you did put so many hours to make it, tweak it so it does the best it could with whatever you have. So, and then suddenly you have that piece of AI coming in and saying, Oh no, we're going to go get like a lot more value out of this building. And it could be very frustrating, to see that, wait a sec. You know, do you know how many hours and years that I put into this?
So, and there's an interesting story about this. I was talking to one of these person that we're describing here, which spent his entire career in one tower. And he's about five years away from retirement, and he's the ultimate expert of that building. And he was looking at me, he said, there was nothing the AI could do better than me. And I was like, okay, yeah, interesting discussion. And I said, well, are you going home at night? And he said, yeah, yeah. Yep. Seven to three, sometimes 3:30. Okay. Perfect. So who's-, if you're not here, what happens? Because there's a shopping mall at that bottom of that tower, during the evening, you know, Thursday, Friday night, that's open until nine. You know, what happens in the shopping mall?
And he said, well, if there's a problem, they're going to call me. Okay. And you like that, to be called? No, no. No. Okay. And what happens Saturday? Shopping mall open, right? Yeah. Yep. They will call me if there's a problem. And what happens if there's something happening during the night and then? Oh, I mean, that's not good. So what about the AI is becoming your second in command, and it's managing the ship 24/7, so you could have a better night to sleep, so you, you could go on vacation two weeks somewhere in Europe. You could go for lunch, not being stressed that your pager is going to ring. And he looked at me and said, okay, now I understand.
And he looked at me and said, and maybe if I show the AI a lot of things in the next five years, maybe I could retire with peace of mind. And I said, yeah.
James Dice: [00:51:26] Yeah. And the building owner would have peace of mind as well.
Jean Simone: [00:51:29] Yeah, so, the AI is not here to replace people. AI Is here to give us the ability to do more on our daily basis. It's increasing our capability or capacity to do a lot more and to maybe focus our creativity on the most important things and letting all of the basic stuff being done by the AI, freeing you up some time to do more important things with your expertise.
James Dice: [00:51:54] Yeah. Fascinating. I don't think I'd be a good passenger in an autonomous vehicle either. I'm an awful passenger as it is. So, cool.
Well as we kind of get to wrapping up here, I want to switch over to the business side. So this is obviously a technology that's on the market. So how are you guys taking it to market? And what's your strategy there?
Jean Simone: [00:52:16] We're focusing on the first wave of building that we test AI at first. So you're talking about retail, you're talking about office tower. So these type of building are, you know, more our focus right now. But we're also stretching into other type of building now, think of data center, airport, hotels, are now becoming a new candidate for the application of this AI. The model that we-, we were trying to basically make it as easy as possible as an offering. So we want it to be a SaaS model, so it's a, it's a monthly fee that you're paying as a service. It's a very, very, very low install cost, so there's basically no downside to go and implement this AI. So you get the box installed, you start to get the training of the AI, and then you pay a monthly fee, which is based on the square feet and the energy intensity you have in your region. So it's a fixed amount of money per square feet you're paying, which is a much lower number than the savings and the AI is going to bring to the table during the same months.
So we want it to be like a very compelling offer where there's really no risk for the building operator or building owner. They will be saving per square feet a lot more than what they're paying on a monthly fee. And the same time is, you know, you don't like it, it's like your cable service at home. You don't like it, well, terminate the service and send us back the box at the end of the month.
So very different than my previous world where we were coming in with an ESCO contract, which is we're going to go in and we're going to basically go to a wedding together where, where, you know, this contract is for the next five, eight years, and it's a very complex contract, and once you sign, you're in for the adventure and there's no backing up. Right? So I mean, we really don't want to offer something like that. So let's make something very, very easy to accept and easy to, basically say yes, value-based, really only value-based. You get the value, you pay. You don't get the value, you don't pay. And the proof is in the pudding. Right?
James Dice: [00:54:26] Hmm. Yeah. So I mean, to kind of summarize, it's cashflow positive, no capital expenses, essentially.
Jean Simone: [00:54:33] Yeah.
James Dice: [00:54:34] Cool. Okay. How about partnering with other vendors? Are you guys going direct to market, or is it channel partners, or how are you guys getting to building owners?
Jean Simone: [00:54:43] Doing a mix of different approach. So yes, directly to the market, have our own sales force, knocking on door, and offering the service to customer. But a big, big, also emphasis on channel, so want to have like the integrator, so the SI in a different region, being able to resell this service. And also having another layer, which is the OEM, so having. control manufacturer embedding this directly as a container or a driver into their controller. So it's AI BrianBox ready. And as that controller is installed, the service is ready to be offered if they want to subscribe to it, but it's already embedded into the controller.
I mean, in a similar type of channel approach, we're finishing up our tridium driver. So, you have a tridium JACE, you just download our driver from the marketplace on your JACE, AX N4 compatible, and voila. I mean, you don't need to install a box in your building. You're ready to go. That JACE with our driver will do our different proximity function and connect to our cloud automatically.
James Dice: [00:55:53] Yeah, and that's one of the things that I think is so exciting here is that, you know, especially coming from my history, which is very similar to your history, is besides like the physical and pride obstacles, there isn't a whole lot of financial obstacle or, really a technology obstacle to getting this hooked up if you think about all the JACEs and, you know, even even in the future with like people white labeling this or installing it on their own supervisory controllers.
Jean Simone: [00:56:21] Yeah, we're focusing on the AI side. For us, our value we're bringing is the AI. All of the rest is the plumbing that is needed to connect and make the data accessible. So, you know, we're very open to any type of a combination that could make sense on the business side.
James Dice: [00:56:36] Cool. Okay. So, you know, it's, May 7th today. And I've been saying that on the last couple of podcasts because, we're obviously in the middle of a very tough time for our industry and for our economy as a whole. How are you seeing that changing buildings and how is it hanging your guys' approach right now?
Jean Simone: [00:56:56] It's interesting because we were talking about that a few weeks back, trying to understand what will be this new world. Last year we were selling more on we will improve the comfort, so, you know, we will reduce issues you have on the comfort side. And a lot of customers were like, it was their key priority. Yes, I want to save money, but I have this tenant on the sixth floor, which keep complaining and we can't stabilize the temperature. And if your AI could fix that, it's worth the money. And there was also a lot of, let's save the planet. If we reduce the energy intensity, I'm going to be able to say that my building consumed less per square. I'm helping the planet. That was really the spin on the sales side.
What we're seeing now is that kind of completely shift now. Now people are, okay, I need to save money. I lost two tenants. They're not paying their bill anymore. My revenue are coming down, so now I need to do whatever I can on the expense line to reduce that one, because the bottom line is suffering like hell.
So we're now becoming like a cost saving tool, and it's probably the only key driver that we need to push forward is you have to perceive us as a no install-, a very low install costs, and I will be able to have a significant impact on that expense line immediately. Well, within the first two months, right? We have to learn about it. But quickly it's going to contribute to that bottom line number, which is now the game, right?
James Dice: [00:58:20] Yeah. Yeah. I mean, I think about how like, you know, it's May 7th. So, you know, at mid March, it was all about how there's no one in the building, right? So can we shut it down completely? Now, as we start to talk about reoccupying, there's people like me that are probably going to be working from home for three more months, I would think. There's not a whole lot of reason for me to go back to the office, but there are quite a few people that are now starting to trickle back in, right? So a solution like this can help keep the energy-, you know, it's not fully shut down, but it can keep things tailored to how many people are actually there.
Jean Simone: [00:58:54] Yeah, I mean, absolutely. I mean, it's-, because of the retraining and the learning, the self-learning, it's adapting quickly to the new configuration. And, and you're right, it's going to be a moving target, right? Some people will go back, not everybody. Take a typical tower and downtown core, probably now it's either completely shut, it's basically on set back 24/7. Or there may be one or two floor because essential services are working on these floor. And then slowly you're going to have this company on another floor starting, but starting with social distancing. So probably half the employees only are going to come back and not the other half, or they're going to do rotation shift or whatever.
So, that control sequence will need to be modulated as a moving target from day to day as this confinement modulation is happening. And suddenly we could have a new research and then the confinement is back quite rapidly. So you need to rechange everything. So we're pretty convinced that the AI will bring a lot of value there. We will continuously trying to reset and basically follow that, that behavior shift, which is happening on a weekly basis. So, you know, once again, it would be feasible to do by sending a control technician to readjust the control sequence on a weekly basis. But I'm not sure how many people will really do that.
James Dice: [01:00:08] Right. Yeah. And it seems like, I mean, a lot of what we're talking about on the HVAC side is it's modulation, right? So, and I think solutions like this allow a building operator to say, how can I also modulate by operating expenses, in accordance with my revenue, which it seems like revenue is uncertain right now. So being able to tailor your expenses as much as possible is, is huge.
Alright, so I also saw you guys just had a round of fundraising, so congratulations there. What's the roadmap look like for you guys over the next year, three years, five years? Where are you guys headed?
Jean Simone: [01:00:46] Well, we're, we're selling the solution. We're already doing several installs on the international side with a few building in Australia and Ireland. We're now going into Southeast Asia. It's quite interesting because even without like the tridium driver, I mean, it's just a box that we could ship by FedEx or UPS, and we're just helping them to install it with a regular FaceTime on our phone. So it's that easy. So it's giving us the ability to, if we have a partner in Bangkok, to just ship them the box. They will install in their own customer in Thailand and, and we do everything remotely. So for us it's quite easy as long as FedEx and UPS are working, we're good. We could ship. So, it's giving us the ability in this confined environment where we could still sell and we could, we could still propogate that, that value generation and different customers throughout the world. So we're definitely going to keep pushing that, keep pushing the development of the things like the driver for tridium that I mentioned.
And then go to, what I call the BrainBox 2.0, which is, okay, imagine if all of these AI engine, which are operating in their own building, they're not aware there's other AI engine working in a building across the street. Right? So imagine if we were connecting them together and they were kind of becoming aware that there is other AI engine deployed another building in the same downtown core. And what happened if they share information, then they start to work as a team to the benefit of the grid. So, you know, if the grid would be like sending signal and these AI engine , knowing what's happening in the next hours in term of prediction could be able to play a strategy to the benefit of the grid, especially if there's a reward allocated to them, by the utility. So we call this the swarm AI, and it's something that we will start working on pretty soon.
James Dice: [01:02:42] Cool. Yeah. It's similar to something I just saw from Google where they're shifting their data center loads across the globe based on time of day, based on how clean the local grids are. So you can think about strategies like that, which are super exciting for the future. Cool.
Well, is there anything else, that we didn't cover that you wanted to sort of say to the folks?
Jean Simone: [01:03:04] No, I think we recovered a lot, and I hope it was not too, too deep, too complicated to understand. Thank you very much for the opportunity.
James Dice: [01:03:12] Yeah, absolutely. I don't think it was too deep for our group. We got, we got some smarty pants that are listening here. So and I'll put my take in the show notes and, again, thank you for, for coming on the show.
Jean Simone: [01:03:25] Hey, it was a pleasure. Thank you very much.
James Dice: [01:03:27] Alright, friends. Thanks for listening to this episode of the Nexus podcast. For more episodes like this and to get the weekly Nexus newsletter, please subscribe at nexus.substack.com. You can find the show notes of this conversation there as well. As always, please reach out on LinkedIn with any thoughts on this episode.
I'd love to hear from you. Have a great day.
Head over to Nexus Connect and see what’s new in the community. Don’t forget to check out the latest member-only events.
Go to Nexus ConnectJoin Nexus Pro and get full access including invite-only member gatherings, access to the community chatroom Nexus Connect, networking opportunities, and deep dive essays.
Sign Up
Nexus Labs exists to move the smart buildings industry forward, without the fluff.
On Wednesday, we send out the renowned (and free) Nexus Newsletter.