#318 Olek Paraska: How AI Is Fixing the Biggest Bottleneck in Construction

0:00Togal itself has tripled in revenue in the last few years. We tripled in revenue three years in a row. Yeah, annually tripling in revenue. To be fair, is that because you're a startup, you're growing from a small base? Is that indicative of how hungry the industry is for a solution like this? I would say the latter. I think the industry is hungry for a solution like this. The industry is hungry for someone to listen. So usually I start by having the guest introduce themselves, you know, explain your background so far as it's relevant. I mean, I imagine you've been doing this for a long time. Yeah.

0:54You've built AI systems across multiple industries. So can you give your background and then I'll start asking some questions. My name is Alexander Paraska, also known as Oleg Paraska. I'm a CTO in Toggle AI right now. And at Toggle AI, we built AI systems for construction industry. But before that, I was building AI systems for end tech, from advertising, where we deployed AI to millions of users before it was as widely distributed. And before that, I was in a streaming media industry where we built a digital rights management system. Completely different angle. You were in building digital what kind of systems?

1:44Digital rights management. So it's just... Oh, that's interesting. Only that people have access to the content. Yeah, yeah. So Togal, can you talk about what's going on in the construction and the housing industry? How Togal addresses it, how AI can apply to it? sure so I stumbled into Togo a bit of a accident but also because I was looking for something like this back in 2020 I think I quit my job back then I decided I needed to start a startup but I didn't know what the startup should be about but I did know that I'm looking for an area where AI would make any meaningful impact the reason why construction is interesting i think is because in the last 50 years the improvement in construction from any sort of innovation has been zero in some cases it has been negative so we have all of this progress in the world in the world of bits but in the world of atoms it's going the different direction and for me as a software person that was interesting and challenging.

3:09I didn't understand why is that happening. And right now it's all coming to the head, but it's something that has been brewing for decades, obviously. But housing crisis is a symptom of that. There's obviously problems with financing, problems with all sorts of things, with zoning. But it also is true that we're just not building fast enough and that october is trying to tackle it it seems trivial but before construction actually starts there's years of negotiations who who does what how do we actually build because you only get to build once and that those years of negotiations and like understanding and planning can be significantly decreased into maybe weeks, months by AI because that is all a very decomposable process where AI can deliver a lot of value.

4:08And this is what Tobel is focusing on. We are focusing on pre-construction. So after the architects have finished designing all of the fancy buildings that they want to build, but before construction starts, this process is ripe for innovation and this is where Togob is on the spot. Yeah, that's interesting. So what kinds of activities pre-construction? For example, when you have a collection of floor plans, this is what typically is being sent from architects. What happens, and I couldn't believe it when I first heard about this, is that every construction person, every subcontractor contractor on general contractor would measure things on those floor plans manually.

4:53So everybody needs to understand, well, how big is this room? Because I need to put tiles in this room. I need to put carpets into this room. And both the tiles person and the carpet person will just go and manually measure the room on the floor plan. And when I say measure, literally the way it looks like is they go and open a file and some program and they trace around the room and that's how they measure. It's not the most efficient process, but for every building, even the smallest building, it happens hundreds of times because there's hundreds of subcontractors, and that seemed like an inefficient process, so this is the process that we're automating.

5:30Instead of you having to draw all of these polygons and lines and count how many doors you have, how many polys you have, we give it to you for free because that's where AI shines, and then you can focus on higher value tasks. Like, you don't want to be tracing things that's what kids love to do you want to design like understand how to build this pattern so stuff like this uh from tracing around from being a perception layer for construct construction but also there's other higher value tasks that now are possible with intelligence like understanding what is the scope here like maybe doing some reasoning it's uh it's ripe for many, many places, but we started from something called takeoff.

6:16You are measuring all of the things on the floor. And so the user gets the drawings from the architect and uploads them to Togal's interface, I guess, and it will immediately count all of the doors and give their dimensions, count all of the rooms. Yeah. Give their dimensions. Yeah, we will immediately, you will upload the floor plans to us, but typically people upload all of the construction documents to us because there's floor plans, but there's accompanying spec books, and there's like hundreds and hundreds of pages of text description of what is it that we want to build. So yes, we are a perception layer.

7:04We look at the construction floor plan. We have proprietary computer vision models that extract the measurements from those components. You have all the lengths of the walls, all of the areas, all of the counts for the doors and windows and stuff like that. That is given to you immediately. But if you want, you can also do this menu as you are doing this before. Because we are sure that AI is AI and you might want to count something differently. And so not only do we give you AI tools, we also give you the regular tools that you have as a manual takeoff estimation. And that hybrid of how much manual, how much AI is always a challenge because AI can do much more.

7:52Maybe you don't want to do a lot of things manual. But our philosophy is always to put you in charge. So you are always in the know. You are always in control. one of our slogans is you are the captain of the ship so we can give the tasks to your mini minions in toggle but you are the captain and you also are responsible for your decisions for us if you misestimate if it's like a huge building if you misestimate the value of the error can be millions and so that's in a way similar to still driving cars industries the value of error can be very, very significant. So that's why there is a balance that you have to strike into how much agency.

8:40Yeah, and the estimating, I mean, I'm also surprised the way you describe it. So they get the architectural drawings, the contractor or the subcontractor, and they literally measure things by hand. I mean, you know, on a computer screen, but by tracing the outline of the room or something, or counting the number of doors they need to, is all of that manual today? Yes, and I don't want to diminish the value of this because in some sense, even being able to do this on a computer screen is already an achievement in some cases because there's also still people who do it on pen and paper. They will print it out, they will take out the ruler, and that's just the process that they follow.

9:36And I'm not saying this in any sort of regulatory form. my point is construction is one of those industries that touches the physical world yeah coming from like an advertising or steaming media it's very very different and you can feel it in the kinds of people who are there for the lack of a better world and i'm sorry i'm using it in the technical philosophical term they are very averse to any sort of bullshit and so because they are dealing with physical realities, they would not buy something that is not going to be helping them. And people say that construction is resistant to technology, but I found that construction is resistant to bad technology, and that is like a majority of this, and this is a big, big problem.

10:24And it's very hard to satisfy a need of a person who's working night and day with their hands, and technology is the last thing that they think about. And that's the kind of difference in construction that you have to work with. Yeah. And so this bottleneck that you're describing, project estimating, you said it's called a takeoff? Takeoff, right. And why is it called a takeoff? Because it's taken off quantities from the floor plan. I see. Okay. And construction productivity has stagnated or even declined for decades. And that is a major opportunity where automation can deliver real results. Was Togol already focused on the construction industry?

11:15Is that something that you saw? Togol started as a spin-off from the largest general construction company in South. I see. And so it comes from a deep knowledge in construction. And the way it started, the founder, Patrick Murphy, he's now a CEO, was looking into his construction company work, his family construction company works. And then he identified this as one of the biggest inefficiencies because it doesn't seem like it makes sense to keep doing this in 2020 or something. And there was definitely a problem to be solved. And AI was brought in as one of the solutions. It wasn't like, hey, we have AI, let's apply it somewhere.

12:01It was completely the opposite. And you quickly saw how computer vision could be harnessed for this task. It started as computer vision. So computer vision is, in a way, a perception layer. So we teach computers to see the floor plans. but now with large language models and generative AI we can connect that perception layer to the reasoning layer that large language models provide and that is increasing in the amount of capabilities significantly because now that the agents can see what's going to work and they can already understand it because they have been trained on construction knowledge in some shape or form, you can take this much further And so there's a lot of back-office tasks that are very automatable because you can connect perception and reasoning.

12:59And that's what we are betting on. So it's not just the vision now, it's like the whole package. Yeah. Yeah. So you've moved beyond the construction estimating into other back-office. Yeah. We moved from perception for construction estimating. We're moving to pre-construction phase mobility. because that phase itself, pre-construction industry itself, is a pretty broad industry. So we're staying in pre-construction. We don't want to go in any other areas, but there is no, for example, system of record for pre-construction, and that's what we are solving for. Yeah, and you started with all of the floor plans by what Patrick Murphy's company was Coastal Construction, is that right?

13:49Correct. And so you started with all of their plans. How do you train a model on this? Do you create synthetic data or you buy architectural plans from public sources? How do you get the training data? So that is a great question because that was the one of the chicken and egg problems that we had to solve early on. So we had a collection of floorplans, but they were just that. They were just floorplans. What we needed to train custom deep learning models, we needed to annotate them. And so we had to find people who would annotate construction floorplans. And it seems trivial, but not anyone can annotate a construction plan.

14:32it's so nuanced and sometimes I wouldn't know what to annotate these things and so definitely it's not like somebody on mechanical turf we had to hire professional architects to annotate things for us and that's a little bit of complexity but in the end we annotated thousands of floor plans we were able to train our first deep learning models for wall detection and for classification and object detection. And then we were able to kickstart the engine and now we have more floorplans and then we can annotate them again. But in this process, we also have produced a way to generate synthetic data because these floorplans are also coming from specific software and there's only so many uses of that software.

15:19So we have a process to synthesize any amount of floorplans and automatically get them annotated. Just frankly, from our experience, synthetic data is always worse than real-world messy data from construction people. At least our machine learning models will always be able to tell what is synthetic, what is not. And that's not a good thing. Yeah, yeah. And this model, is it an open-source model that you're fine-tuning, or did you build a model from scratch? We built a model from scratch. there's multiple models but some of them are transformer based so in a way it's inspired by a transformer architecture but it's completely trained from scratch by us some of them are some object detection ones like UOLV6, UOLV7 I think and so we took those models and then we went to them on construction floorplans but for the main segmentation we have trained from scratch yeah and so is that a smaller model i was just talking to a company cohere do you know them they're i'm actually have been very involved in them yes yeah and so they build small models that people can that people can deploy on site is that important for what you guys do or are you a cloud-based model that they hit for inference via API or something?

16:52First of all, it's a small model in today's terms. It fits in PHP view, but it's a pretty big model in terms of how deep learning is concerned. But we do not deploy it on-premises. They hit it, people who use it through an API or through our front-end service, which is a web application that we provide. There's many reasons why deploying those models on-premise is not a good idea, but not the least of them is in construction, people do not want to mess with complexity. They don't want to combat complexity where it's not needed. Why go through the paint or host in the model if you can just use it through an API?

17:36Yeah. And you're moving into agentic AI. How does that work in construction? Specific in pre-construction, pre-construction is almost all back office. So it's all operating with some sort of documents and moving documents around. For example, one of the cases is you as a subcontractor for drywall. You receive a set of documents, but it's not clear for you, am I supposed to do this or that? and so like it's a very typical process that you receive a document and then you need to send an rfi for what what is it like we want to request more information but to generate it rfi it's like it's a whole process you need to generate it because you need to log it for future purposes because later if you didn't ask for it or you didn't clarify you didn't have the logs you are responsible.

18:32That whole process in between, it's like you generate an RFI from a text. It's very routine. There's nothing that should not be automatable here. Stuff like that is very susceptible to automation of using agents. Another example is, again, you receive a document, but you just don't understand what's the scope for you. What are you bidding on? And there's like 500 pages of documentation. Some of them is relevant for you. Some of them you just don't care about. You're sitting in a truck somewhere in the field. You just want to say, what am I looking at? And so just parsing all that information, summarizing it for you is a perfect fit for an agent.

19:13But we come through and what our core strength is you can do all of these things using large language models. What we bring in is we add a perception layer because now these large language models can understand the floor plan. And so from there, you can just say, here is one version of this floor plan. Here is another version of floor plan. What are the differences? What are they meaningfully different? And compare them analytically and just tell me, like, what do I need to ask? What am I not even thinking about? Big part of pre-construction is analyzing construction. Sure, architects might have thought about this, but you are a construction person.

19:52You know what they don't know. So large language models have that baked in, but you need to tap into it in a specific way. And that's another kind of agency, that agent that can shine in. Yeah. And you mentioned sitting in a truck. Is this a field-centric solution or are you sitting at the construction company headquarters figuring this stuff out? Typically, you would sit in a construction company headquarters. It's a back office. One of the complaints that our salespeople have is that when people join from their trucks, it's very hard for them to sell them because they have a million things going on.

20:32That happens, but it happens only occasionally. You would typically do your job for pre-construction in the back office. I see. And these agents you're talking about, they can handle things like identifying materials or managing workloads autonomously. Can you talk, just walk us through a use case that would demonstrate some of that? Okay, so for example, you have a floor plan, you are a drywall contractor, a subcontractor. We talked about this, for example. what toggle gives you out of the box is just a line for where the wall should be. And you know that drywall probably should be a drywall.

21:16But the next obvious step for you as a subcontractor, you need to understand through this line for the wall, there's a small tag connected to it. That small tag says something like M4, M5. And you as a subcontractor know that this means this is a specific type of a wall and I have to go and look through all of the other pages, like 200 to 1 ,000 pages, and find where is that tag defined to understand what is this wall. And you find it in like a page 275. This wall is a specific kind of a drywall. It has this many cornerbies, this many studs, and it has to have two sheets of drywall and stuff like that.

21:57Now that you know this, the next step for you is, okay, I need to connect what I have discovered here to this quantity that they have here. And that's a non-trivial task right now because, again, you have to do this gymnastics. And from there, you need to understand, well, when you are estimating how much would it cost for you to build this, it's not just the quantity. You need to understand what are the materials that will go into this. So you need to build all of the assembly for the wall and count every stud, count every sheet of drywall. Then it becomes like an optimization problem that is, again, susceptible to computer thinking and cognition, but it needs to be framed that way.

22:41And if you can go from this process from the very beginning, you had nothing, to the end where you have a bill of materials that you may need, this is the ultimate holy grail. And that's the collection of agents that we are building. And so the vision is that a contractor can take the architectural drawings, give them to Togal's system, and then query Togal through natural language of what's the total cost of this project in today's market. and it would go through and figure out how much steel girders you need, how much glass, float glass you need, whatever. And it'll spit out a number, a cost number.

23:35And then can you then say, do the agents, can they order the materials? Can they say, you're going to need X square feet of drywall. Let me order that for you from a supplier. Or how granular or active does the agent cut? Yeah. Do not have plans to build agents that will order for now. There is different plans that will probably be better suited for that. Perhaps a marketplace where people can bid on the project if they know it would be better suited. We don't try to go there because that's a bit of a different phase of construction. Again, we try to stay in pre-construction. On the other hand, if you ask about vision, the vision is that when you can map from a geometry of an architect to a bill of materials in the real world, in the physical world, that is a very powerful thing to have.

24:33Because when you have that, you can quickly change the plan and understand how it influences the value of the building. And that's a lot of things that can be solved by value engineering like that. So for example, for housing crisis, okay, you cannot afford this much. If we remove these things, this is how much it would cost. Ultimate vision for Toggle will be like you have a prompt, and that prompt will generate a plan for you, and then plan will generate a bill of materials. So sure, we might be able to generate those plans, but we still want to have humans involved for other reasons. So the real value is no matter what the client is, we can give you the materials.

25:19And then if we generate plants, if architects design plants, it becomes less relevant. Yeah. Yeah. And then the contractor or the subcontractor who, as you said, knows, has a wealth of knowledge in what works or what doesn't work or what's easier or more difficult. So can they then ask the system, is there a substitute for this material which is not available right now? Or is it that kind of thing where you can have a conversation with the system about the building? Exactly, yes. So now that the system knows what is it talking about, in a way it has an internal representation of BIM. like a bin building information modeling without you explicitly having to build bin because that's an expensive process on its own.

26:16So it builds this bin internally and then you can query it. Some of the questions people would ask is if I want to build this building but make it more green, give me some recommendations. What would make this building more green? Is it maybe a different kind of concrete or like order concrete from here or maybe insulated for, I don't know, for other reasons. So you can then reason on top of this because you have all this information already there and the agent is already connected to it. What kind of questions you can ask is up to you. Typically, we see people ask questions about constructability again because that's important.

26:54If an architect forgot to have a drain for a shower, it's nothing for architect, but it's on you because you're responsible. And then you have to tear down the walls and reconnect it. Like, what am I missing is typically what people want to know, because that's something you cannot even predict. What am I not asking? Other questions, like I said, what is the scope for me? Give me some recommendations. How can I decrease the costs? If I want to decrease the costs, how to make it more luxury? Those are the kinds of things that you can get into once you have this internal being representation. Yeah.

27:32Does it, I mean, it certainly, I would guess, increases productivity, but does it, I mean, beyond the time savings, does it lead to real ROI on a project because you've been able to avoid problems, as you said, see whether or not there are drains in the showers, or you've been able to adopt cheaper building materials? because this model makes it easy to test what would work and what wouldn't work. Where is the value for the contractor? For the contractor, the immediate value. The construction company. Yeah, for both. For both of them, the immediate value is you would have had to spend a day coloring polygons.

28:26You don't have to do it now. So you just save the time. That's immediate value. The second value is, other than just having always fancy AI, it's probably not as important to talk about this in a podcast like this, but you are bringing these floorplans on the web, which most of these construction companies do not have. And now you can collaborate between different estimators. And so that collaboration process is already increasing the efficiency because you don't have to send those files in PDFs, and email us one to another. That itself is a huge value gain. From AI, you just press a button and then you have to understand what other quantities there.

29:10It's also a huge value added. The chat that is connected to the application is something that is being developed right now. So I can give you concrete values for ROI. We know that people use it for quickly, quickly decreasing time for understanding the scope, answering the questions. They just ask a question, what is this thing? And then they have a response immediately. So that's where we are right now. And although there's a lot of stuff to build still. Yeah. But again, are there some metrics on how much companies can save or how much inefficiency the system can save So going from a day to an hour of estimating time.

30:02So that's typically 90 % efficiency saving on estimating. Yeah. Why is construction productivity lower now than it was in the past? I can give you theories, of course, and I don't think I'm well qualified for this. Just coming from software, that was also a question I had. My understanding is that construction, actually a lot like software, is a lot more about people systems and less about technology. And in my experience, software is also a lot more about people systems. But construction, people systems combined with physical realities makes it super hard to adopt any new challenges because the value of entry is super high.

30:53You have to understand construction extremely well to be able to deliver value at least a little bit. And frankly, that's also what you're feeling in Tovo, the company. We have people who come from construction, and this is the first software that they are working in, the software company that they are working in. And we have people coming from software, and this is the first construction company that they are working in. And there are these just not understandings of each other inside the company. And we're actively working on trying to figure out how do we talk to each other? Because the worlds are so different.

31:28And the way we solve this within the company is also how we solve this for the industry. But in my mind, there is such a big gap between the technology and construction that it's a gap that very few companies can jump through. Why is construction not improving in efficiency? I think, again, because it's not about technology as much. It has not been about technology as much. It was more about people's system. And I believe why right now we are at a point where we can really change things because AI is the only thing that can quickly unravel all of the backlog of innovation that hasn't been there.

32:14And it really democratizes what is possible, but without the gatekeeping of software people. And this is why I'm excited to work on a project like this. Yeah. Yeah, it's fascinating. It's an industry that has not, certainly it's changed with technology, but it doesn't change quickly. And you're working on massive buildings, I would guess. Yes. Primarily, we work on commercial buildings. So that's like big buildings, typically it's hotels, hospitals, and stuff like that. There are some construction companies in one of the customers, among the customers of Torval, who are building like family, one to two family homes.