AI Can Now Estimate the Entire Job, with Dmitry from Handoff의 전체 녹취록

the latter one. So testing is something that is very important, you know, if you're building any kind of AI product. We actually, you know, ⁓ over course of the years, we train our model in thousands of blueprints and we build our own benchmarks that show the accuracy for our product versus any other foundation model, let's say. and we actually made those benchmarks open source. And so if you want to validate the accuracy of your let's say own project against those benchmarks that's available. So we for the sake of the takeoff bench, yeah. So we set a very, very high bar for what we would put into production against those benchmarks that are quite hardcore. And we actually saw that the model performs better than human estimators. So this is just a quality bar piece that we wanted to reach before it actually becomes available to everyone. And on the other hand, you know, the way that the product is built is the thing about pre construction, and I think you guys would ⁓ you know understand that is Luigi La Corte: This is takeoff bench. Dmitry: Pre-construction is also there's a ground truth to it, like a takeoff has a ground truth to it, but it's also dependent on the way that you bid your jobs, the way you organize your jobs, the way that you price your jobs. And we give our customers power to first of all see the outputs, be able to then talk to the model to then customize those outputs, and to be able to provide in you specific requirements, like well, you want to break out your estimates floor by floor, trade by trade, right? That is considered to be a portion of what accuracy means in construction. Luigi La Corte: Mm. Dmitry: Right, but this is a subjective piece of estimating that you still need to incorporate if you want to do a good job. Well, we through the product we give our customers the power to be able to do that as well. Luigi La Corte: It's ⁓ it's pretty. Mike Powers: So I guess it's like to go yeah, to go one level deep like player deeper on that. Like what if it is wrong? Are you guys taking on like the risk of being wrong since it's not kinda wanna say closed source, but end to end? Dmitry: Yeah, I mean like like any provider of software, you know, you you know, we provide the tools and software that is needed do a takeoff, right? We we we allow you to do it faster. We give you all the outputs so you can go and validate the outputs before you actually go and submit this as a bid. so it's works in any, you know, in the same way that your software works, right, Mike. And so I think it it's maybe about the same about the same approach. Mike Powers: Inter yeah, I I misunderstood. I thought you said like meant like it was like a black box and it outputted the answers. so it is still like trust but verify. So it's like here's the widgets on the the takeoff screw the the PDF or whatever. Dmitry: Yeah, you you get a full you have three outputs. You get a full markup. So all the all the files, all the markups of the files, we see them represented on the on the plan. You get a scope report so it actually goes in detail about all the things that are going into the project, like all the scope of the project and all the kind of s you know, specific details that we've been able to derive. And because, you know, we work with architects and great you know, we all love architects and you know sometimes they just don't do a good enough job on a set of plans and that's okay. and we we do try to capture those either errors, conflicts, assumptions, and we create a transparent list of everything else that we saw that might be a risk that the customer should be aware of, or any assumptions that we had to make. And so they can make a judgment based on those. And then the final piece is the estimate with the takeoff included and all the pricing connects to our construction supplier partners, like we work closely with Lowe's, Home Depot, ABC Supply and others. ⁓ and so all the pricing can be traced back to their catalogs too. So if you wanted to see well, why is this priced in such a you know, in such a way, you can see exactly where it is in the c your fervice supplier's catalog. Luigi La Corte: Mm. It is a ⁓ it's amazing. we'd love to can we see it? Dmitry: Yeah, yeah, I'll be happy to ⁓ click around. Let me see, how do I share this? Luigi La Corte: Let's do it. There should be a button fifth from the left that says share. Dmitry: Awesome. Okay. ⁓ yeah, so this is the handout platform. ⁓ so it it you know as as as as is a is a kind of an end-to-end construction management platform. It does multiple things. You can manage your contact book, your CRM, your construction schedules through this, do your invoicing. ⁓ and this is kind of your pipeline, right? If you have a bunch of products, pro bunch of projects that are in the works. So this is a ⁓ a takeoff that we just did. ⁓ on a single family residential home. I wouldn't go into like all the details that you can do with the platform. So we can just go right into the takeoff here. But this is, let me just kind of give you guys an idea. So this is an output, right? It says shows all of the trades here. And then as you kind of expand this, you see the breakdown of all of the trades. And as I was mentioning, the prices, you can trace them back. ⁓ You know, since we have enter here, look at let's look at joists. So ⁓ You can trace them back to the you know Home Depot in this case catalog. You can purchase actually through us as well. ⁓ now when it comes to the takeoff, you guys can see, you know, we uploaded the takeoff, we said, hey, very simple. Here's a plan, go do the thing, right? Don't have to overthink it. And then it ran for it, was a set of agents doing the work on this, using our ⁓ computer vision models, using our data to put together this takeoff. And this is the steps that it went through. So, like if you know, as the takeoff goes through the process. It took us about one hour and fifteen minutes to do that. I would say a typical

estimator would probably take, I don't know, ⁓ a few days, right? To put together a full takeoff with an estimate. ⁓ if that's all that they're focused on. And this is the outputs that you get. You get the ⁓ the estimate right here, you get the drawings and get a summary report. So let's just jump into the fun part and look at the drawings together. ⁓ these are the drawings that you get from ⁓ from from from handoff. And all of this was done. ⁓ automatically. There's actually no way, at at least at this point, to to draw anything in a handoff by hand. And so you can see it goes into such detail as deriving the ceiling joists and making sure that it's done against the ⁓ also the kind of the engineering constraints of a given project, right? Because the joists have to be placed against ⁓ you know, the the kind of the the walls that can reinforce them. ⁓ and then it goes through all of the other details of Luigi La Corte: Mm. Dmitry: All of the other scopes, like exterior siding, ⁓ your doors and windows, ⁓ your roofing, and then your electrical plan as well. And so everything is accounted for. If you guys want to see, we actually have this option to replay the stream of the model doing the work. ⁓ so we can can show you guys how that worked. ⁓ and this this should go at ⁓ at 5x the speed. But so I'm not touching the computer right now. This is literally what the model does when it does the takeoff. Luigi La Corte: Yeah. Let's do it. Dmitry: It goes through every plan like an estimator do ⁓ would do. It evaluates the details, it evaluates the call outs, the ⁓ you know, any tables, and then it starts to segment out the plan to what we call the roys, so you know, regions of interest. And off of these regions of interest, it be you know, it builds this sort of understanding of the job, right? And then maybe I can just jump a little bit ahead. And again, this is what an estimator would do, and this is why it's such a tough problem. You cannot just do a takeoff, like, you know, it's not just a bunch of drawings, it's really the understanding of the entire project context. And so it starts by kind of measuring, creating these different details. Often enough, they kind of look like nonsense because the model doesn't work the same way as human does. But as it goes through that, it starts to clean it up and it starts to kind of like populate it one piece by piece. Right, and evaluating its own work and doing its own ⁓ you know, counting and it's doing its own quantities. So let me see if I can show you like a cool visual while it just like populates things. Let's see. Maybe let's just do it at a page scale as well. So it doesn't jump around so much. So you can see over here the model just goes through and it just populates those ⁓ those those areas and it just adds the takeoffs, adds the markups as it goes along. From these markups, it then derives the quantities. ⁓ and there's multiple really difficult problems we had to solve in order to do that. Right. It's like, well, like how do you accurately like distinguish between interior and exterior walls? ⁓ you know, ⁓ left handle door versus right handle door, you know, different types of windows. sometimes the plans don't have everything that you need, right? Sometimes the let's say, ⁓ Luigi La Corte: Mm-hmm. Dmitry: You know, sometimes you might just have a floor plan and like you have to work with that and you have to make assumptions about roofing, you have to make assumptions about all the other trades. Over here is kinda cool. You can see is drawing the the the joist. I I thought that's like really neat. You can clearly like you really see how it's happening in real time. ⁓ Luigi La Corte: Yeah, that's so cool. How can I ask a specif a very specific question? ⁓ typical details. Typical details ⁓ not always scope. ⁓ scope if the plans if the plan context requires. ⁓ how are how does the model specifically choose to implement ⁓ let's say the the w what would potentially impact quantities if a typical detail is relevant? Dmitry: ⁓ yeah, I mean look so so it it it does so different plants have different level of detail that is available. So this is like a what, like ⁓ twelve page plan or give or take, right? So ⁓ you know, where where it does where the details are available, this is what it's using, right? It can actually look at the details provided by the architect or by an engineer, and then the what's in the detail will be referenced and then used to then make the estimate and the takeoff. Now, when it is not available, we've built a massive ⁓ ⁓ you know, called a knowledge base of construction principles. We started the company essentially doing estimates and takeoffs for people by hand. And over the course of that initial period we worked with hundreds of estimators and with them we've been able to obtain all of this very detailed construction knowledge of how to actually build something. And so when a model is trying to reference this this this sort of like library of construction knowledge, ⁓ when it doesn't have all the details available on the plants, this is where it goes to, to be able to then Get those answers. Luigi La Corte: Yeah, very very, very magical. so w let's talk about ⁓ performance relative to benchmark. So you have takeoff bench, which handoff introduced. what are what are you actually reporting on relative to takeoff bench? Is it is it mean average percentage error for the objects that are being drawn? Or does it depend on the task? Dmitry: ⁓ yeah, so ⁓ what the takeoff bench does, it analyzes performance on a set of specific trades using the accuracy of what a a perfect ⁓ human estimate would look like in terms of quantities, in terms of let's say units of measure and scope completeness. And so we down score ⁓ the outputs against those trades. and then you're able to then quickly we and we use this evaluation harness called Harbor, and you can quickly plug in a model within that evaluation harness against those scores on

the set of permission blueprints that we've ⁓ derived from our platform. And then you can see how your model performs against what a perfect estimate would look like. And what I mean by perfect is when we did that in the beginning with just human estimators. We actually found that human estimators, even the ones that are quite experienced, there were a lot of imperfections. and so ⁓ then we had to go through multiple rounds of revisions to really get it to like, okay, nothing, absolutely nothing is missed to able to set that bar. And then we were able to compare their their their their kind of like the the best possible early submission against the final submission where all of the revisions were actually accommodated. Andrew Zukoski: Yeah. So does this i is this primarily residential or do you do commercial scopes as well? Dmitry: Yeah, we started mainly in residential and so I think where it excels the most is remodel, single family residential, but at this point we also do commercial. So some of the other types of projects we do cover quite well are ⁓ retail, whether that's new construction or TI, in a multifamily. you know, I would say w where we do still you know, have not explored so far is you know, larger projects, maybe more civil projects for sure. ⁓ but kind of the bread and butter of the majority of construction here in the US, the model covers pretty well. Luigi La Corte: Hm. In ⁓ yeah, go go ahead, Andrew. I'll ask, I'll ask after. Andrew Zukoski: Okay, and then I was gonna ask, like what happens to I'm a I'm a million dollar year, two million dollar year Resi builder or remodeler. what to my if particularly adopt? the estimating piece. Do I estimate more? I estimate better? Do get to like not build shopping lists and have to go to Home Depot? I just get to hit buy and then all the parts show show up on site? Like what happened like why is my life better? Dmitry: Yeah, fantastic question, right? So let's just maybe take a step back and kind of thinking through what a typical you know, life of ⁓ somebody who runs a construction business like that looks like, right? So because, you know, they may be running a a a business that doesn't have a ton of staff, they have to do a lot of the work themselves, right? And so that's being on on the job side, you know, actually managing the projects or managing your team that is growing, doing recruiting. So like, you know, we're all founders here, right? So these folks are also founders in their own way. And so wearing a lot of hats. Now, how do you make money in construction? Is through bidding, right? Primarily, right? You have to bid and we have to win those bids in order to actually bring in the cash flow. and so I would say there's two things that kind of are a benefit to using AI to doing your estimating, and you know, applies to anything, whether that is through handoff or not. ⁓ it is one is you can do more work, like you can actually cover more bids, because you may have more opportunities to bid than. than than you currently can physically accommodate, respond to, now all of a sudden that expands. ⁓ and the second is the quality of your bids is better, right? Because now you're trying to grow, you're trying to bid more. Every single bid, know, it is again like it it it is how you make money. The profitability matters. But the more you try to do yourself by hand, the less accurate you're gonna be, because again, simply not enough time. And so you kind of get back both, you know, the higher accuracy, higher profitability, and more revenue on top. Andrew Zukoski: Do the my impression don't know very much about the home building space. We briefly considered doing a renovation and then got estimates back and that was the end of that, right? Like but my Mike Powers: Ha ha Andrew Zukoski: Like the estim we went through this process. Like it was like a hundred-year-old house with like some real geotechnical issues, like it may not be typical. but my impression was there wasn't a lot of accuracy in the estimates were being brought back to me. The like and I don't know if we were like there was just like a huge I mean this was a substantial renovation, there was just a huge amount that was unknown. Right, like we had a drawing set, but even then, like no one had taken the walls open, like we didn't know what was gonna be under the soil. Like, how much does the accuracy matter in residential like why why does it matter that I've got like a more more accurate, a more precise estimate in the space? ⁓ Dmitry: ⁓ you mean ⁓ matter for you as the client of the contractor or matter for the contractor? Andrew Zukoski: the contractor. Obviously I want complete price certainty and I want fewer digits on the price. ⁓ yeah. Dmitry: Totally. Yeah, I mean it's just because ⁓ it's a project based business, so it's all about the inputs into your project, right? And so and then typical, you know, construction firm tries to hit a you know, and there's a whole like argument of, you know, markup versus margin. But you know, a typical construction firm is trying to do thirty percent mark ⁓ markup, right? Let's say, right? And they forget to include all the other expenses that they need to incur, you know, to actually make their margin. But so like if you don't estimate correctly in the beginning, you know Let's say lumber prices went up, right? And you just didn't incorporate that because you're referring to lumber prices from like a year ago. and so if you don't do that, then your you know, markup ⁓ right? And the resulting margin shrinks. Now, what you're referring to also is like the on-site conditions. And when when a contractor does an estimate, usually in their contract, they say the estimate is based on like the best understanding of the job that I have today, right? And so, but what often happens, especially with remodels. You know, you probably open the wall and you find out that like, I don't know, like your plumbing is from the nineteen

fifties. And like we have to redo that too. and so in those scenarios, usually the contractor would change or would do change order to then accommodate ⁓ those things that they just did not know at the time. and then the good news is that you can actually do the change orders through handoff as well. Mike Powers: I guess one like well one level deeper on that as well. Like and again, I'm not from the Redzi space, but like for that for the example like Joyce, like you kind of know the spacing that's needed in a Joyce open. I'm sure the estimators that a home builder would know that. And they know the square footage of a space. Isn't kind of just a simple calculation to go, hey, it's you know, a three thousand square foot first floor, thousand square foot times whatever the calculation to get the linear feet of Joyce that I need. Andrew Zukoski: Right. Mike Powers: Is it does it actually need like this detailed of a takeoff them? 'Cause like I'm guessing not every home has like really detailed drawings. Like what like that there seems to be have to be some sort of balance between two. I would guess. ⁓ Luigi La Corte: Hm. Dmitry: Yeah, i yeah, definitely a great question. I think there's multiple layers or multiple types of estimates, right, that you might ⁓ that you might do as a contractor based on the type of work you do. and so and also like the price of the project. When you are, let's say, a remodeler and you're doing like a twenty K bathroom remodel, the so so so first the complexity of the project is lower. Right. So you're actually not gonna most likely have any joys there, but like you still wanna quantify things like amount of tile or fixtures that you're gonna install. and so in that scenario, just this is like a simpler estimate. And for the most part, it doesn't even need a blueprint. Like you can, you know, ⁓ inside, take some measurements, take some photos, upload them into handoff, describe the project, and you can be done and you know to a high degree of accuracy. now as you go up, you know, then the next question you have to ask yourself is You know, are you let's say you're doing a single family home ground up? Now you have to ask yourself, are you gonna be taking on any of that scope yourself? Right? Are you self-performing or you're contracting? Right. And if you're stop contracting here again, like you can actually even start at like you know design development phase with like a pretty simple set of drawings where you don't need to quantify every single joist, every single stud, right? Every single door handle. is going to be good enough for you to kind of go with like a more high-level estimate. then now you can sub out to other, you know, ⁓ traits that would then perform ⁓ the work. And then the most detailed, the most hardcore is you're framer. and and actually you're gonna go and frame frame that house. and this is where, yeah, 100% you actually do have to quantify them or you're a roofer, right? You have to quantify everything, it has to be detailed. Basically, any trade contractor, right, they will need to have that in-depth level of detail. This is where drawings come in, and this is where detailed takeoffs come in. And so we can scale across all three possible scenarios that may find yourself in. And and and and this is the reason why not only the model works really well for the contractors that are using the product, but actually offer the model to suppliers. And for suppliers, the accuracy just has to be like really on the dot because they're actually gonna sell the specific to the contractor. right. So the suppliers who use our model, ⁓ they actually use that to quote the actual, you know, number of studs, number of joists that it will have to be selling to the contractor ultimately. Luigi La Corte: Yeah. Andrew Zukoski: Who takes the risk then? Like if the supply like A I don't know if ABC is an example, customer of yours that does this, but if it I I'm A I'm ABC Supply, I work with a bunch of home builders, I offer this value-add service, I say like yeah, throw the throw the plans over, I'll put together you're gonna need. You buy it, you take it on site, you deliver it. Like if there's a miss there, who owns it? Dmitry: Yeah, well hopefully everybody does a good job, right, in that process, right? Both, you know, with if it's handoff providing the takeoff, right? And ⁓ and the and the associate quoting the the job and then the contractor taking or you know, taking a look over the quote. But in a in an event that all three miss it, it's really the exact same process if you would have a human estimator, right? Let's say you you're gonna you come in, you know, with a set of plans, and say some suppliers outsource that to remote. Andrew Zukoski: Mm-hmm. Dmitry: Locations, right? Where the the takeoff is done on their behalf, they get the takeoff back, let's say some somewhere in Southeast Asia, right? And then and then they get the takeoff, they give it to the to the client. So that the I I would say that the 80% of the time, this is how actually it's done today in the lumber and building material industry. you know, if if there are, you know, ⁓ you know, God forbid issues and errors, usually there's a return process. ⁓ right, where you can then let's say, Hey, you gave me too much, you know, too many studs here. Like, you know, I I you need to go pick it up. And so it's a fairly common process in the industry where the re Andrew Zukoski: It does seem like you're probably always gonna be over rather than under if the supplier like does the estimate, right? Like you're probably not gonna show up to the site without things. ⁓ Dmitry: I mean, yeah, and ideally you got the right measurements too, right? So you don't have to cut it again. yeah. So but ⁓ but but I would say kind of like the returns, delivery and returns process is a part of kind of today's world and lumber and material material ⁓ lumber building materials industry. But again, like this is

where I think we ⁓ just all of us, right, in construction tech and create value is just grinding down all of the situations where the errors do arise and improving our products and improving our models. Because ultimately, like if you do ship your takeoffs off to you know a remote takeoff service, there's only so much you can do to improve it there. Like it really just gonna depend on how well the person running, you know, that that specific location is gonna be able to accommodate it. And I think that's really cool about what we do here is that we get to control and learn and constantly iterate. And ultimately I think this is why we can provide us kind of like as a construction tech, I would say space overall, provide ultimately a much better service. than than than anybody else would. Luigi La Corte: We we see like a similar well the implication of your question, at least in how people use our product, where like we're not perfect, so you you be can you're called a checklist. At least our products are different. Like we we do scoping, but not the quantity takeoffs. Quantity is just they're just very, very hard in commercial projects. I guess two part question. One, how are using it? Do they use it as a as a check on their existing work? and then two. ⁓ ⁓ l let's say let's say there's they're saying, okay, I'm actually using this to be my estimator, which does make sense because if you trust your estimator and they do a good job, you are not auditing their work. You're simply taking it at face value and say, okay, you know, Dimitri does a great job. So yeah, it's probably within five percent accuracy. I can live with that. You know, I'm gonna I'm gonna move on. But let's say let's say it's working very well and people are replacing it. Is is estimating solved, right? Or are there specific edge cases that still need additional data? Or, you know, some kind of novel approach. Like where is the where's the challenge? 'Cause like electrical and sprinkler, that's hard. That's not on the drawings. That requires understanding of how conduits are run and fire sprinklers just show you the location but not the pipe run. So very long but what are your thoughts? Dmitry: Yeah, no, it's a it's a it's a it's a perfect question. I think that estimating has two components. It always has a ground truth component and it has a judgment component, And and every bid, like every estimate is ultimately becoming a bid, right? And every bid has a component of like, what do we want to do here? Do we wanna win this job, right? And maybe bid less. Do we wanna, you know, just gonna like this client, we'll bid you know, we'll bid as much as you know, as high as possible. You know, hoping that they, you know, reject it. and so I think from the standpoint of like what is ground truth, like that is, I would say, it depends on what what what solved means, but I would say that is largely solved because if you do provide information into the model, right, the model will then make very good decisions with that information. So if you have a blueprint, that's a lot of data that you can base your takeoff from, you can base your estimate from. And so the kind of the ground truth component, the ground truth accuracy is gonna be there. you can also provide a description of the project, right? And then in that case, it really just depends on how good your description is, right? And ⁓ same thing you I'll come to you and you're my estimator, right? And and and I'll tell you, hey, we're gonna do a you know, this hundred square foot bathroom remodel. And then you'll be like, All right, ⁓ is there anything else I should know? Right. So I think there's like a lot of detail that needs to come into that. You can be a perfect estimator. But whatever you give me is just going to be based on the amount of detail that I provide you with, right? And so like the takeoff is a good one because there's as much detail as you can derive. So I would say it's sold from that standpoint. Ultimately, I think every human, every every every owner of a construction company, even every estimator who's using this, they need to add their own spin on the output in the same way that let's say you go to Claude and you say, Hey Claude, you know, write me an email. You probably will, at the end of the day, the email will be correct, be sendable. You can send it exactly as it is, but usually you'd wanna, you know, kinda tweak the last, you know, last mile of what that of what that's gonna look like. Luigi La Corte: so if you were to aim H at a commercial set, I know you said you started dabbling, would your benchmark that you built, the takeoff bench, does it encapsulate every discipline, you know, depending on which CSI you're using, divisions one through sixteen, if it's older or up to forty something if it's newer, does it do all that very well? Or are there some divisions where it just doesn't work yet? Dmitry: Yeah, great point. So and just to clarify, the the takeoff bench one, so this is the first version of the benchmark, right now contains only residential projects. ⁓ we have not added the commercial projects in the publicly available benchmark. Our internal benchmark does include commercial projects. ⁓ and we'll be working on takeoff bench two ⁓ to then include the commercial as well, make those publicly available. Now, as far as the CSI, like let's say go trade by trade, I I I need to understand why, but for whatever reason, we actually beat Essentially, we were the best in class on every trade except for insulation. why insulation? I do not know. ⁓ I actually need to we're trying investigate and understand it a little bit better. But other than insulation, like we're pretty I think we're like we're pretty locked in. Luigi La Corte: That's amazing. Yeah, that's that's that's remarkable. Dmitry: Yeah, and it was insulation was and when I say installation it was only like a couple of percentage points off. but we're like, Okay, I guess we haven't spent

enough time on insulation yet, so we'll we'll have to you we'll have to we'll have to delve on that a little bit more. Luigi La Corte: Yeah, change the temperature from zero to one and then maybe it'll Andrew Zukoski: Yeah. Dmitry: Yeah. Luigi La Corte: there's this like really I like very annoying discourse in construction tech about it's like pretty protectionist about ⁓ you have to be from the industry. Mike asked this question in the chat. your what what's your view? I mean obviously ⁓ we can infer what your what your view is, but why do you think it's not relevant? Or maybe we just don't understand your background and you can inform us as to how you came from the industry. Dmitry: I mean it's a g I mean ⁓ it is a question that actually would you know, I'm definitely having to take my give my take on it. Are you still not from the industry? Like at what point is there a cutoff? Like ⁓ Luigi La Corte: Yeah yeah. Andrew Zukoski: the the the year of your life, Mike, that you spent drilling micropiles in the like Eagle Stadium, Luigi La Corte: Ha ha ha. Mike Powers: Jeez Christ. to say, just to state, I think over the last two months I've drastically shifted my opinion on this. I was definitely on the camp that you had to be from construction. And I the tech companies that are coming out right now, the founders are not from construction. outside of provision, of course, because I love provision and Luigi. ⁓ and I think people are gatekeeping. ⁓ so for example, what what Hollis all started was ramp release, like ramp for construction. Luigi La Corte: ⁓ thanks. Mike Powers: And people are like bombarding it like, hey, they have no idea what they're talking about. They're not from construction. I'm like, Ramp is a credit card processing company that does expense reports. the same for a contractor as it is for an accounting firm. ⁓ gonna crush it in the market. And just like started this process of like, are have I actually been gatekeeping people being like, ⁓ no, I don't trust this person because they're not from industry. so I've drastically changed my opinion on that. And think a lot of the cold tech is coming from people outside the industry now. Which is awesome. 'Cause like it's it's like open up the tent, Like we need to stop being like you spend six months sitting in the snow of a micropile, ⁓ like micropiles. Like ⁓ it useless. None of that's re referenceable to what I'm doing in in construction tech right now. So Andrew Zukoski: Yeah. Luigi La Corte: It's only relevant to the extent like you're not successful as a result of that. Like if you can't talk to a customer because you don't, you know, ⁓ l the language, then yeah, maybe it does matter. But Andrew Zukoski: Yeah, but the like the construction industry I don't know, this my I came from the outside, right? My experience is that th the construction industry is collection of loosely related industries that you have to you have learn how to speak to a great diversity of audiences to navigate most problems of scale in the industry, right? Like Luigi La Corte: Yeah. I think I I I agree I agree. I just think like the way we talk to investors bleeds over into how you market your company and you feel like there needs to be some like fatalism. Like if you didn't swing a hammer at twelve, then there's no way you could build a global VC backable construction tech business. It doesn't really matter. but I agree, like if you're yeah, if you're working in the industry for seven years from tech or, ⁓ you on site, like yeah, you're you are from the industry. If you show up day one And expect it to be a forty million AR, yeah, it probably matters, but y you're not there. Dmitry: Mean I Mike Powers: Just talking about the coolest ones. You guys know Mick. He's not from construction, doing really cool things. Like I mean, Bobbyard is probably the best takeoff tool right now in the space with training road models. it's I I just Luigi La Corte: Bob Yard. They don't they don't get the recognition. Don't you find that interesting? Like the people who talk about the best tools, like Bob Yard's very rarely mentioned, but they I I mean it's hard to infer, but they seem like the fastest growing. And Michael's team is quite quite cracked. Quite quite cracked. Yeah, extra commercial. ⁓ Mike Powers: Commercial. Yeah, for commercial. So it's interesting. Those three the the the between you three, it's it's an interesting data point. Is causation correlation? Dmitry: I think I think that I think that it's a blessing and a curse for what we do, ⁓ like kind of all of us here, is that construction is not an intuitive place. It's it has a lot of things when you're like, Well that that is weird, like why is that? Right. And you kinda have to be very curious and you have to sort of play bowl with how people do things and you have to come into this with lot of respect. and I think it takes time to get there. And I think if let's say we were starting, you know, like I'm a you I'm a machine learning guy, like I can maybe I can do like a DevTools company, right? if I did that, I could probably would have like you'd know answers much sooner, you can relate to the people that you work with much quicker. And the error rate ⁓ the lowest. And I think in the world of startups, venture back startups, you live and die by your error rate because that is what correlates to your attraction and your runway. And if your error rate is higher than what your runway allows you, right, to to to sort of afford. Then you go out of business and then you become, you know, the story that everybody would be like, well, you know, was Katara, like nobody was from construction. They're like, you know, I mean, we can probably d you know, discuss Katara. ⁓ Yeah. probably a bit a bit a bit bigger, right? But but it's like you have to and if you're like a founder listening to this podcast right now, Andrew Zukoski: That probably wasn't the only problem there. Mike Powers: Ha ha. Dmitry:

Construction tech is amazing. I love people who like I love contractors. Like I I love hanging out with contractors, but you're gonna be like the dumbest person like ever for the first probably, I don't know, two years, three years, right? Until you go through all of the stupid mistakes you're gonna make. Because you just your assumptions coming from tech do not correlate well to the assumptions of people or to the way that things are done, of people that are in one of these pockets of construction, right? And if you go Like we sell into contractors and we sell into suppliers. There is some overlap. I mean, sure, like there's similarities, but very, very different go-to-market motions, very different needs, very different requirements. Even within the suppliers, like different suppliers are like a big box supplier versus like a hardcore, like you know, local ⁓ L BM supplier. And so you just have to accept that there's gonna be you're gonna be an idiot for a long time. but if you're okay with that and you can like survive and convince your investors that eventually you learn, it's probably a good place to be. Mike Powers: So I don't I don't envy or go to market at all. Curious like how is that I mean there's there's probably one two guy shops in cities that you can sell to and then like large enterprise accounts on the residential side. What does that look like? Have you guys like scaled by regions or how are you thinking about getting in front of all these customers that are pretty sparsely populated, sparsely spread out? Dmitry: Yeah, totally. I mean, what we do here is not to dissimilar from, let's say, company like Toast, right? Toast that sells into restaurants that are like of, you know, small to mid sized ⁓ type locations. and you know, the the nice thing about the S and B market is there's a couple of things that are nice and there's a couple of things that are not nice. So the th just starting with the things that are nice is there's a lot of them, right? I think within our target market there's maybe eight hundred thousand firms, right? So we have plenty of room to grow and like I think with You know, maybe if you capture even, you know, let's say 80,000 of those, like we that'll be like would be at a company of an IPO scale, right? So there's a lot of room to grow. and then when you sell to SMBs, if you position your product well, you can be not just like a point solution, but you can be their own, you know, their full OS, right? You can run their whole thing for them, which I think is much harder if you go into commercial construction where there is more complexity. You know, ProCore obviously succeeded at covering the whole thing end to end. Luigi La Corte: Mm. Dmitry: But I think it's a lot of people already have ProCore, right? So like, well, you have to go and compete with that. and so but being their OS, being their end-to-end solution that really is where you can monetize a lot better and you can have a higher share of wallet, better retention, is easier in SMB for those reasons. Sales cycles are shorter, the decision makers on the phone, you know, when you call them, it's easy to find them, they're local businesses, a lot of really good stuff. Right? Well, things that are a little bit more complicated is like you may call them and they're maybe in the on the roof, right? ⁓ And they have to make a purchasing decision from there. There was a guy I was talking to and ⁓ it was it was wild. Like we were we were we're doing a demo, he's like, Yeah, I'm gonna take the coal from the truck, no problem, right? So a lot of time people take actually their calls from the truck. Guy comes in, his face is white, completely white. He's covered in paint. Literally the guy just painted, you know, painted a house. And I was like, Are you sure you wanna do this? Like, we don't have to. Like maybe he was like, No, no, no, this is my every day of my life is like this. Andrew Zukoski: Yeah. Dmitry: So you sell into people's like, yeah, you sell into customers that are that are in these in this in this and I'll say often like in challenging positions. And so it it is it is it is a different sales process and you kinda you have to be you have to change your mind about how you sell into SMB and it's about velocity, it's about clear clarity in your value proposition, brand really matters, you have to be super product led, the product has to be intuitive. There's it it's a it's a hardcore task. to be able to reach that market, but if you do reach that market, I think it's a really awesome market to be able to sell to. Andrew Zukoski: I mean it I assume Toast handles all the financial ⁓ and so I'm I'm guessing they have factoring products ⁓ top of that and they they start to start very meaningfully impact ⁓ the they interact with. Is that where Is where handoff goes, becomes sort of the back office in a box that brings all these these value added services to home and to smaller contractors Dmitry: Yeah, that's exactly what we are, right? So we are an agent that helps you run your business. We say AI teammate. and you can delegate work across the entire back office. And so you kind of get both a worker, your AI agent, and the tool that the worker can use. So you you know as a part of handout platform you have budgeting features, scheduling features, project management features that are tightly connected to the agent. Right. So it's an entire back office that runs the firm, you know, the the company for you. And payments is definitely one of those ⁓ elements. Custom we actually have best in class payments ⁓ rails that we build fastest payouts, lowest fees. and then the ⁓ c contractors love using it because they can accept payments digitally through us, whether that's through ACH or through credit card, and so then connects back to their bank account. And there's a variety of other fintech products we plan on launching in the next future, not in your future. But it's payments is definitely like like

a toast or, you know, other kind of companies in the vertical workflow space is is typically where you go to. Andrew Zukoski: Well and so why do both suppliers small contractors like why why work with those two very different audiences where it sound it sounds like a huge part of building like there's obviously some very exciting technology that's been built. We saw a demo of it, and there I'm sure there's more as well. but huge piece of building that business and bringing bringing the technology to market is the distribution and the support for the builders. Like, why take on a very different set of problems, I imagine, around big box suppliers. Dmitry: Question. Yeah, I think it's there's ⁓ two two parts of it. It's like one is by choice and one is not by choice, I would say. the part is by choice is that you know we've we've spent we believe that if solve these problems like cost estimating and takeoff, you actually solve the fundamental problems of how to construction. Because a takeoff is basically what it is: it's a cheat code into the rest of the project. Like you can quantify ⁓ All of the materials you need to buy, you can quantify the labor you're gonna need, you can quantify the time it's gonna take to complete, you can connect that to your purchasing, you connect that to your team management, right? So if you solve that piece really well and you automate it, you kind of earn your way into becoming an bit of an infrastructure layer for the rest of the pr project, the company, and ultimately, because construction is very interconnected, right? And if ⁓ you know ⁓ GC works with a sub or GC works with a supplier. All often enough is the same piece of data that gets sort of transferred from one sort of pocket to another. And you can unify everybody in one single data standard, right? So you can become that infrastructure product, infrastructure Rails. and so we decided to sell into this specific segment of general contractors because feel like it's the best place to start if you want to build something like that. and suppliers is always something that we wanted to do. It was always a part of our vision. And we started working with them in a different capacity in terms of like bringing their pricing into our platform. But then some of them started coming to us and were like, hey, like you guys seem to be doing well in these like you know quoting and takeoff things. Can you do it for us? And ⁓ you know, we were just very lucky because they basically gave us an opportunity to work with them and there's one, you know, ⁓ one big name that has us implemented in almost two thousand, ⁓ two thousand of stores across the country. and so when when you know as a founder when you have an opportunity like that it's hard to say no. you know, we probably like we wouldn't be seeking that out ⁓ you know, out of our own accord at this point in our company stage as because we kinda had an opportunity to work with suppliers that became another market that we ended up tapping into. Andrew Zukoski: That's awesome. Luigi La Corte: Dimitri, so ⁓ on like w on one hand, I'm I'm surprised that the market pulled up pulled this out of you. maybe not on the supplier side, but on the on the residential contractor side, because it it feels like it's the minority of cases where they have a full set of plans to actually take off. it's usually like, here's what I think, take some photos, take some measurements, ⁓ you know, do that. So maybe you've already answered that question, but how did you land here on takeoffs on these drawings? seems like more of like a commercial construction thing. And and you've mentioned that you're going there. Andrew Zukoski: like our commercial c construction customers would say that it's a minority of the times that they have a full and complete set of drawings to do takeoffs from. Like Luigi La Corte: But you think you think the proportion ⁓ to no drawings? Okay. Andrew Zukoski: No, no, no. I don't think the proportion. I think it's I think it's gotta be more acute the residue. But Demetri would know better than I do. Yeah. Luigi La Corte: Right. Yeah. So yeah, love love your comment. Yeah, go ahead. Dmitry: Yeah, but by the way, though if you guys ever want to integrate H1 into your platforms that are serving commercial construction, we would be very much happy to do so. And so you you know you guys can kind of enable those capabilities there, just putting it out there. as far as as as far as kind of like ⁓ why take off, it depends on the market segment. ⁓ home builders they operate exclusively off of blueprints, right? It's very, very common. So if you sell into home builders, custom home builders, you will have an architect that has designed a home and then you have to build off of those plants. now, and that's with still within kind of like the same residential construction segment. often enough you have remodelers that are trying to break into new home construction or commercial TI, for example, that often comes with plants as well. So the demand in our market is definitely there. Now, but when it comes to building an estimate, there's sort of three components to it. There is the there's the quantities, there's scope, and there's the pricing, right? And so we actually started with takeoff back in 2019. You know, a little history on me. I was I was working in this company called Adams, it's Travis Collins company that used to be mainly in cloud kitchen space, like dark kitchen space. Now it does all kinds of things. And I was trying to figure out how do we use machine learning to automate the process of real estate development of these kitchens. And so this is where I learned about takeoff. And so I was like, well, this is like sounds like a perfect problem for machine learning to solve. And so we started working on takeoff at the time, but we discovered that actually it wasn't the right first problem to solve. And so what we focused on instead was pricing and scopes. And so the

first version of handoff platform that we launched did not have takeoff. And you can do all of your estimating through descriptions, photos, and so all the non-takeoff, like non-blueprint related types of inputs. Once we felt that the model was perfected enough, this is when we enabled takeoff. And so for our home builder customers, like I mentioned, our customers that cross over into con commercial construction, that has been like a really, really nice addition that they've been using. Andrew Zukoski: Mm. Luigi La Corte: Yeah, that that makes sense. man, I was gonna ask you a second question. D do you guys have anything else you wanna ask while I while I try to remember where I was going? Micro on you. Mike Powers: I ⁓ one one other question was I know like the word model gets thrown around a lot. are you guys leveraging like a foundational model and then doing post-training? Are you guys using an open source model with open weights? Like what is what does look like under the hood? Dmitry: Yeah, the word model is a bit of a like a nebulous term, right? Because it can mean a lot of different things, right? I think models traditionally means is just something that, you know, is a com is is is a combination of components, right? It can be like a combination of weights, right? It can be like a something that's simple, it can be something that has like a bunch of different layers to it. Like a most, you know, basic regression model is just like a formula, right, with a couple of variables, or you can have something that's incredibly complicated. so in our case, where we shine, I would say, is we've trained our own computer vision models, and that has been done on like thousands and thousands of blueprints, with you know hundreds of estimators that worked with us. And so those computer vision models that are forming the basis of the accuracy. Right. On top of those, we actually use agents. and those are the agents that we ⁓ you know, we we rely on foundation models and the AI labs to offer access to those agents, but those agents operate within our harness. ⁓ And within the knowledge bases that we've created, and they have a certain workflow that they go through in order to accommodate the work. So there's a combination of different pieces to this that allows us to do what we do essentially. But the core, you know, the models themselves, ⁓ the the LLMs themselves, or ⁓ you know, even with vision capability, they're very weak when it comes to working with blueprints. So, but they're strong in other ways like understanding context or using tools, for example. Mike Powers: So not launching Kim Kimmy three point tomorrow on the back, not not spinning up a server in your office, come on. Dmitry: That'd be so funny. No, ⁓ you know, actually I have an announcement, you know, H one has been banned by the US government. ⁓ it's Mike Powers: only people I want to van are is now OpenAI and Anthropic, which is hilarious because they're gonna see their market cap drop, but that's beside the point. Luigi La Corte: Yeah. Dimitri, I am ⁓ so look my dad's a general contractor and he doesn't do custom custom builds, but he did a lot of work for for clients. is the question I was gonna ask. Are you also doing conceptual stuff? So homeowner says, Here's a picture of my here's a picture of my bathroom, here are the finishes I want. I don't have a design. Can you give me a range? how well does H1 do on that? I know you probably don't have a benchmark, but what do you what's like the vibe on on how that performs? Dmitry: Yeah, H one is a model that is specific to takeoffs and that's the one that we've launched this week. But for the the pretty much the entire kind of time that handoff was live, we would operate more of these like you described, maybe conceptual or description based prompts. And so ⁓ actually a lot of consumers use handoff. And so if you guys ever do a remodel and you wanna double check, actually one of our investors she's she's ⁓ she's doing a remodel and she checked it against handoff and she was like getting ripped off and she brought it back with her contractor. So we kinda we help her save a little bit of money there. but no, this is actually the I would say where we started and we're spent a lot of our time at the beginning, just based on photos, descriptions, videos, ⁓ you can you can do that as well. guys, might have time for one more question. ⁓ Luigi La Corte: Mm. No, no, that's it. We're we're we're we're gonna wrap. I just I just wanted to say, very impressive this I was texting the guys in the group, this is a magical product and I'm super amped to hear how the industry react to this. If I mean if you could maybe tell us very quickly how how they are reacting. ⁓ but this is definitely the future. So congratulations on H one. we're very amped for you. We're obviously in your corner ⁓ cheering you on, this is gonna be fantastic for the industry. Andrew Zukoski: We gotta wrap. Dmitry: No, likewise, guys. I mean, it's such an honor truly like to be able to talk to you about it. ⁓ because I think that in construction tag there's always like a lot of skepticism. There's a lot of like, you know, like, ⁓ is this just like another hype thing or something like that? Right. So we we want it to be as transparent as and open as possible. That's why we're publishing a paper, that's why we have a benchmark, that's why the product is available for use. that's, you know, you know, it's it's it's a real thing that we've been working for a very, very long time. And ⁓ ⁓ the fact that, you know, to hear from you guys being experts in construction tech, it really means a lot. Luigi La Corte: Amazing. Congratulations again and best of luck with the with the implementation and rollout. Mike, you're on mute if you want to say something. ⁓ I mean experts is a stretch for sure. Maybe except for maybe except for Andrew. Mike Powers: ⁓ sorry, they say experts is a stretch at construction tech with these three with talking heads around the circle. Come on our pod. Andrew Zukoski: Three dipshits here. Yeah. Wait, they're experts? Where do we talk to them? Dmitry: I said in in construction tech. Con construction construction tech, not construction. Luigi La Corte: Alright everyone, we're signing off. All right, that's right. Alright everyone, thanks so much, Dimitri. Thank you, Andrew, and I got a jump. ⁓ we'll talk soon. See you guys. Bye. Mike Powers: Yeah, yeah. Andrew Zukoski: Even that. Dmitry: Thanks guys. Bye bye. Mike Powers: CS.