رونوشت کامل Takeoffs in One Click, with Elan from Kamai

to America, ⁓ right, you have money. We don't. So we we just to figure it things out. And this is how we ⁓ essentially dominate across many, many fields. Not not a lot of people, we are you know, wise, and money wise, not a lot, but we are very determined to win and stuff. That's that's the culture, essentially. And this is and also from a a from a financial perspective and geopolitical perspective, we don't have like as a state, as a nation, we don't have really enough resources. And like eighty percent or ninety of our export is really about technology. Mike Powers: Love it. Love it. Elan: So we don't have commodities like oil or energy or or anything like that. just export technology. ⁓ Mike Powers: Very cool. love it. yeah, so one of the questions I wanted to ask you today, ⁓ kind of like the big news. so there's kind of two big news stories this week in construction and also in in LLMs in general. I want to start kind of on the LLM side. So there's obviously been a lot of traffic and conversation around like open weights, closed weights. you had a really good position that we kind of chatted through on Monday about like your thoughts on foundation models. would love to give you yeah, kind of get your thoughts and opinions on like like This conversation matters to us as technology providers, but does the customer actually care if the models open weights, close weights? what what are your thoughts on that and how are you positioning that to customers? Elan: Customers don't really care about open or close or whatever. essentially they're all cost money. If it's open or closed, someone has to pay pay the cloud bill, right? To to ⁓ for example this Kimi K ⁓ three that recently was launched, you would need a huge cluster essentially to run it over, and make it ⁓ make good inference. Not only that you would need to also manage inference. ⁓ From a tech perspective not ⁓ as easy as it sounds, right? So if you have on your phone or on your computer, a lot of things are in the background running the actual model. So a for a vendor who needs for a for a for a service provider like a consultant company or an estimator or an engineer you would essentially you're essentially for output for outcome not for the actual like you care what's going on inside and behind the scenes so I don't think it really matters. ⁓ of fact we are of course every software company codes with ⁓ clawed code ⁓ you're doing specific classified work that you cannot really share information outside of your closed environment you won't be you never be using weights models because you're they're just not ⁓ as good as the closed Mike Powers: Interesting. So are you not seeing so have you like done any testing with I know we're kinda going out a rabbit hole, like with Kimmy three's model that do you think it's not as consistent as like Opa ⁓ Opus or Fable or Seoul? Elan: I actually haven't tried Kimi K three. Andrew Zukoski: Yeah. Luigi La Corte: Kimmy. Yeah, Kimmy came out. Well the the

latest model came was very recent. Yeah, it was like two two two, three weeks ago. It was only it was only Frontier on front end. so it's not like it was Frontier in any other like terminal bench it wasn't. Andrew Zukoski: Okay. Elan: I try I tried the Kimi two point four, think, models ⁓ they just not as good as the ⁓ as Opus. ⁓ So ⁓ Luigi La Corte: Sure. Yeah. It's Kim it's Kimmy Three which made a huge leap. Elan: I'll be definitely checking it out. Luigi La Corte: Yeah. It I think irrespective, it's like, yeah, the ⁓ like the customers clearly don't care. It will benefit them though, right? Because if someone who's gonna avoid paying margin to open AI and Claude is just, you know, yeah, you're right. They're gonna be they're gonna be paying for compute still. Like they'll have GPUs or whatever. they're gonna have to pay for those things. But it's but it's great. Like I think ⁓ I think it's fantastic. I would love more open weight models that allow me to diversify beyond Open AI and Cloud and and Google. That'd be great. I still think go like each model's still very good though. Elan: I think it generally speaking Yeah, definitely. I think generally speaking ⁓ the the open open source race is really advancing everybody together, right? If ⁓ open AR AI or clouds are lobbying the AI market right now, if release an open source model, it is much better than their closed source model. ⁓ It incentivizes them to release more and more and better models and doing more stuff, right? So it's an accelerator to the market. This is my ⁓ personal opinion about it. So more research, more open stuff will just advance the ⁓ industry in general. Mike Powers: I think one like one funny data point was ⁓ obviously like Anthropic's like raising their hand saying ⁓ they distilled on our model. but didn't anthropic distill on the internet slash all the books that they've kind of illegally processed? And it just seems like it's like another step down in the in the journey. one comparison I heard, which I don't know how real which I made, so it probably is a bad comparison, is it like is it illegal for someone To try to match the flavor profile of Coca-Cola by ⁓ the chemical compounds and doing distillation of chemicals out of Coca-Cola. It would be illegal for them to say, Hey, here's our Coca-Cola brand and here's our logo. That's trademarkable and IPable. But is the actual output taste ⁓ it and IPable? And it seems like weights are kind of like a taste in that sense. ⁓ Andrew Zukoski: But it might be in it's probably in the terms of service. Like we have in our terms of service that like you're not gonna reverse engineer our products based on using it and like you have to agree to that to get in. So it's not like it seems like it would be a civil thing, not a criminal thing, but ⁓ Luigi La Corte: Yeah, that's what it's what it comes down to. It's clearly a terms of service with at least with Claude and you know and Moonshot AI. But Elon, like for for Kamaya, obviously I I I don't know the pricing, ⁓ but I've spent some time on the website. It's very impressive. I I I'd love to get into some of the accuracy and to what takeoffs have not been solved, but it sounds like hey, upload your drawings, you can get full quantities. so like objects, lint la linear lengths. ⁓ area areas, object counts, ⁓ volumes, I'm not sure. but that's hard that's Mike Powers: Yeah. W before he goes into that, I would love to like applaud him on also having like the idea of an open AVAI open API the and it's literally self startup, yeah, or self starting in the application. So I didn't ha I to Elon just like Andrew Zukoski: You guys didn't go in, I got a key. I ran a bunch of drawings. Mike Powers: I don't keep no, I was like saying I I talked to Alan just because I was like interested in chatting with him. And then I went off and like literally embedded it in our application, tried the like tried the entire process. It was super smooth, super fast. he has a really, really cool way of like kind of capturing geometric shape and then finding those replicates, the replications across the application. so I applaud you on that because I think a lot of their like specifically takeoff founders have like buried this like, hey, we have this like super secret sauce that no one's allowed to use. yes, but love to kind of have you like talk through like, hey, like what what was your thought around your go to market motion? Why did you like why did you decide to like lean into the open API first and your your kind of embedding ⁓ ⁓ ⁓ the iframe in the application? would love to kind of be unpack that. Elan: That's a question. Well, guess a year ago, ⁓ much, we saw ⁓ ⁓ even even before we started the company, we saw the other competitors in that space and what they're doing. ⁓ trying to sell their own products and force users to go on their platform and use their platform to do stuff. now ⁓ specifically in people are not ⁓ t tolerant towards Right. People are It's a physical world. they're dealing with physical things, and the one thing they hate is software. But they to, right? They have ⁓ manage their invoices, they have to manage their staff, they have to do takeoffs, they have to place bids, they have to answer emails. so it's unavoidable. ⁓ So, how can we it from a product perspective as easy as possible for And this is like the big LLM revolution ⁓ that has been going around here, is you can speak to the and the instance you ⁓ can tell the computer hey I want this and that and it actually does it So that's that's the big leap ⁓ that's is ⁓ happening and fascinating, amazing. ⁓ You can your it started even before that with Siri, for example. Siri play music, Siri call ⁓ someone ⁓ and it actually it, right? The LLM part the actual more intelligent ⁓ layer is an to it. where saw Kamai is really we deterministically a specific question that a pain in the ass those consortium companies, right? We want to know the exact measurement and the

exact accounts of for example ⁓ walls doors rooms ⁓ da da da da ⁓ to make their and ⁓ place their So how do you make it our to market was really okay, ⁓ we make a product and essentially have a funnel B to B funnel where we just schedule a bunch of book ⁓ calls with a bunch of customers ⁓ and we a close rate ⁓ and this how we grow, ⁓ which is standard. But we are AI which we do, and we develop niche and specific models for this industry, how do we make it ⁓ ⁓ how do we make most progress plus most profit in the short in the shortest period of time, right? And figured we can do that with other software companies. So we literally just were developing our models and our ways to interact with other softwares. this is how we got started. this is we got great design partners to ⁓ partner up with, and ⁓ we got to real user feedback on a scale. And we right now we operate on scale of ⁓ 4,000 something users that are using our software. And it's been amazing about it. And the adoption also. Actually, had a conversation with a client yesterday. He bought ⁓ ten pages on our ⁓ on my website ⁓ and realized he needs more, so he bought the bigger subscription and he started it through cloud. We've opened our MCP in our MCP we also released yesterday a beta version of of a user interface. ⁓ So essentially Kamai lives inside your cloud. can also is also released on Open AI. Kamai is also l ⁓ Kamai also Lives in your open AI Chat GPT interface. you once plug in Kamai into your ChatGPT or Claude, you can just Hey, what's going on? My project? Bam! You get accurate information, you get the actual print, we get the actual markups, yeah, this is like from go-to-market ⁓ We are not forcing our users to our platform. We do have a pretty cool platform, ⁓ it's pretty refined and and nice to but ⁓ we're not you to do anything you cloud. or OpenAI or I got an email from ⁓ Autodesk they have released their MCP connectors as to ⁓ allow you to upload the your MCP onto their ⁓ website. If you're any of those platforms, we easily can within and yeah ⁓ that's our growth so far. Mike Powers: Yeah. I love it. I think it's like it's it's th I'm like a super fan of what you're doing. Kind of like I think it's like under like you're still like underselling the entire like there's so many things that you can do with the tool out of the box right now. ⁓ know you've had like really, really good experience not just on like selling to software companies, but also selling to general contractors and like helping them set up like, hey, you already have a cloud subscription, you already have an open I a open AI subscription. ⁓ ⁓ when you need to do something that Claude can't do or open AI can can't do, is ping into this like API, this MCP and like validate, get the accurate. Here's the actual dimension of the wall because it's a geometric model as opposed to an L L ⁓ when you talk to customers, like how do you bifurcate those conversations? Are you very much like, Hey, like, LMs are really good at X, Y, and Z, we're really good at A B C and and how do you like kind of educate the customer market on that since there's there's obviously s there's so much noise right now. In the space of, hey, my proprietary model is the best thing in the world, or you can do it in open AI. I'd love to kind of hear how you think about that. Elan: ⁓ so LMs are you know language ⁓ large language models. They're mostly based around language. The the models that ⁓ usually people are using ⁓ cloud or chat GPT or those or Grok or any any of those kind of models are multi like ⁓ vision, language, also a bunch of other tools that they incorporate behind the scenes. ⁓ For example, if use Gemini, ⁓ behind the scenes use ⁓ document AI, which is one of their cloud services to analyze ⁓ documents. Now they do have like a scale of ⁓ how they parse your documents and you also adjust it. But essentially it's they can interpret user commands and instructions towards other ⁓ This what what is going on the scenes. You can think of Kamai as one of those tools, If have an intelligence layer that is exponentially growing and getting better and better, like Kimi, like Outpus 5, 5, maybe Fable 6 in a few ⁓ months, right? this this awesome, this is but you would need from a construction perspective deterministic output, would need to know the exact measurements of like for example, you're ordering doors, right? you would need to know the exact width of a door ⁓ in order to like actually put it inside the the the otherwise you're millions you can't really you can't really a mistake here this is where we in we give you exact measurements and is ⁓ what the user they really need another wrapper or another interface another ⁓ place salter that they need to go to they just want the answers and fast ⁓ so ⁓ yeah Andrew Zukoski: So but then like the ⁓ I agree on people wanting the answers fast, but a proud developer of workflow software, I feel compelled to defend our honor, even if I'm totally wrong. you need to do something like there's a few things. You need to be able to do something with the answers. Often you need to able to understand the answers, and then you have to be able to correct them. Right. Like you have to be able to go in there and understand what they look like and correct them and so forth. And so doesn't that doesn't that point to like quite a bit of like I'm not it's not just like chat in, chat out. Like I I need something that lets me in and interrogate the interpretation of the drawings and then I maybe this is where you're saying the things like Autodesk MCP servers allow you to push the data into whatever the downstream like tool or workflow is. Elan: Yeah. Yeah, so this is exactly why we partner with software vendors. and plus we have our own product which is for Takeo specifically. So you can visualize and link the actual data towards what's going on on the planet. And you can validate and adjust the parameters. is also what I said about the MCP UI. So MCP

allows you to also incorporate a user interface ⁓ inside ⁓ thread. and I can ⁓ maybe show it ⁓ a little later how it looks like right now in the beta version. ⁓ But essentially, when ask ⁓ Kamai, hey, what's going on in my project and show me the markups and ⁓ and everything else and measurements, it'll actually pop you up a pop-up your thread and show you okay, this is the blue. Luigi La Corte: Okay. Elan: Print this is those are the markups. is what you're referring to. Then can validate, say yes, no, adjust, and so So you have the visual validation of information. And the point is, for example, in coding, yeah. If we go back in time, as a software vendor, you tried obviously ChatGPT and Cloud back in the day when it was just released, ⁓ right? You're writing code and then you're checking up and Luigi La Corte: Mm. Elan: Checking things and if they're working, if it wrote it correctly, and so on, so on. You still do it now, but do it way less, ⁓ Because you trust it more, because it actually performs and gets you the right answer. I been or any other developer on my team. we're just writing code anymore. We're speaking to the to Cloud and we're validating, okay, this this can be this can go to production, ⁓ but the the the go to ⁓ the time to POC and the time to production has decreased successfully. significantly this because the models are getting better and better exponentially so is what we are betting on once ⁓ users trusting enough in the system they can check it's fine and they should again it's millions on the line ⁓ but This is ⁓ where we're aiming to a one shot one shot way of a user to actually query our API, I need information about this blueprint, about this project, bam you get it. you do with it it's fine. You can use our you can use ChatGPT, you can use Cloud, you can use Build Vision for that matter, you can use any other tool for your specific niche, ⁓ because we're building a very f low of information and Then you would need to build up your workflows. For BuildVision, for example, we spoke I spoke with Mike on Monday, and doing a software essentially for vendors ⁓ who sell who sell mechanical equipment, ⁓ right, Mike? they need certain specific piece of information for my software, for example, the mechanical takeoffs, and other ⁓ vendors who doing invoices or purchase orders, their software is all about that. They need another thing. So everybody takes little piece. ⁓ Luigi La Corte: Mm. Elan: utilizes it and amplifies it in their software. So is like the end product. Mike Powers: You wanna show it? You can you can show my account if you want. ⁓ if you want to just like walk through I think a visual B is super helpful of this. I think the coolest thing was like again, the fact that like it instantaneously set up, like it instantaneously runs. There's no like pre training that needs to be done. he like they have not spent any time on mechanical systems, so this was kinda like a net new exposure. He's done a lot of with like

doors for so we kind of just like went in raw and said, Hey, like let's see how close this is. so kind of just like setting that expectation. Andrew Zukoski: Mm-hmm. Luigi La Corte: This was Mike with work that you've or like you already had an eval that you were testing this against? Elan: Yeah, ⁓ Mike Powers: So and what's interesting about it is like of the challenges that we have is like LMs are good at like reading the the tags. Actually, can pull it. Like reading the tags on a mechanical drawing. but getting the exact location of like the the the grill register diffusers is challenging for an LM, like with bounding boxes. so we were able to like say, hey, like look for this specific shape, and it recognizes it pretty quickly. you know, within the 10 seconds on all the drawings and counts all of those shapes automatically for us. So in our world, we're like, hey, we're using an LLM, we're able to like ⁓ the find what we need on the drawings. We have context around what we need. And then this is that step of like, all right, now let's actually make a perfect bounding box on the piece of equipment so the user knows exactly where it is on the drawing and can trust but verify it. and the ease of use of like if something is missed, actually I'll just show you because it's like if you go to If something is missed of course it's not gonna let me share. ⁓ it's not gonna be ⁓ Riverside does not let me share. can any Alan, can you sh it's gonna kick me out. I'll be right back. B R B Luigi La Corte: Sure. Maybe maybe while Mike's getting that out, ⁓ Elon, how how does Kamae think about Is it you know is it like overlapping area, like is it ⁓ A P E? Like w how do you think about it and and and how do you guys perform relative to benchmarks? Elan: there are not that many benchmarks out there simply because to benchmark correctly you would need actual ground truth ⁓ and no has actually released an ⁓ an ground on actual production ⁓ documents. ⁓ I haven't anything that is like remotely close to what we are ⁓ benchmarking against ⁓ in our own evals. ⁓ And that ⁓ Luigi La Corte: Just ⁓ coincidentally, last week we had ⁓ we had someone on for handoff, ⁓ Dmitry Alexin. They just published a takeoff benchmark apparent apparently. It's called ⁓ takeoff bench. But that's I think a week old. Yeah. Elan: awesome. A we a week old is already old enough. but but yeah. ⁓ maybe it's pr something recent. definitely check it out. We are going to release open source also our own benchmarks to actually look and compare other softwares against ours. ⁓ Not Chat GPT but also our competition right. There are just a few companies that are doing that stuff we're already performing better than they do. Mike, you're presenting my product? That's awesome. Ha ha Mike Powers: Sorry about that. It's just brutal. Brutal movement. all right, cool. Let me Share this real quick. ⁓ so what I thought was really, really powerful can you guys see my screen? So what's like really really po powerful about this ⁓ actually go into other one. ⁓ that's fine. ⁓ so if I go in and say like let me find all of which ones have I not done yet? ⁓ these E ones. Luigi La Corte: Mm-hmm. Mike Powers: So for build vision, we already have all like the context beforehand of this. So like we know that there's E1s on the drawings. We know there's how many there are in the drawings. ⁓ we still are not able to grab the shape of the E1. So I can actually come in here and we're looking at ways to just automate this. and it's gonna actually go find all of these shapes. Elan: You can zoom you can zoom in if you want to. Yeah. Mike Powers: like pretty quickly, ⁓ see if it is right. ⁓ pretty quickly goes and finds all of those ⁓ the drawing and the exact location of where they are. So like that's like the big unlock for us ⁓ the accuracy around the location. ⁓ so you can see here, yeah, grab these two right there and tag those. So now ⁓ have the exact location of the drawing of where those boxes are. 'Cause on an engineer set of drawings, like we have all this information already. So we have what the shape looks like. I can feed that in automatically. We have how many there are. and our value is like, hey, can we actually go do that? ⁓ and you can exceed, you can see I like I don't really care about like the linear feet of this pipe, but it's an easy capture of actually going to find the linear feet of this pipe because they've like done walls and doors and frames and hardware. So like, hey, I'm any type of construction tech company or any type of user and I need to go get like linear takeoff ⁓ It's it's like as accurate as it can be 'cause it's it's G it's not like guessing numbers, it's not using an L ⁓ It's literally like measuring a line pixel by pixel in the drawing set. Elan: ⁓ well not pixel by pixel. ⁓ as I said, geometrical models are different ⁓ a little different. There is not much research about it. ⁓ I think we're Mike Powers: Not things like yeah. Elan: Perhaps the only company in the world that is doing actual production for those kind of models. it's kind of novel ⁓ approach towards ⁓ this problem. And inherently cat drawings geometries, ⁓ right? Somebody wrote it in a software ⁓ that and have the precise properties of the vector graphics in that. So our are we Pretty big models on predicting geometric properties in to predict the actual shape. And ⁓ it comes down to ⁓ Luigi you asked about accuracy. In our world like it's a it's a binary world. It's a zero or one. Is it true or There is no in the middle. Andrew, yeah? ⁓ Andrew Zukoski: Mike, can you pull back up that the same thing we were just looking at? Mike Powers: Yeah, yeah, yeah. The share screen is brutal on Riverside. ⁓ my gosh. Andrew Zukoski: Yeah, fuck Riverside. I mean I love you, Riverside. Stop making me look like a NL. Mike Powers: I just can't like you guys go into this like weird box that's like vertical every time I hit share screen, so I can't even see you guys. Yeah, go

ahead. Andrew Zukoski: Yeah. Okay. Okay. So I can't draw on this, but I'm this just has me curious. I'd love to love to trace through like how a geometrical model handles this differently in a previews. Can you put back on the objects that you found? Like make the E one seventy five thing appear? 'Cause it missed it missed one. Like there's the two up there that it had. Mike Powers: It's yeah, 'cause one thing it's looking for. So it here's an example. Like it's looking it's looking for this exact shape. So you can see that like, hey, like this this is actually a different shape, this one down here. So that's why it Yeah, so do this. ⁓ the other one? Andrew Zukoski: Yeah, that's that's one it missed. Yeah. Elan: Go ahead, mark mark mark mark it up. Show it. No no no not that one. The other yeah. Mark the other E one seventy five. Andrew Zukoski: So I I My understanding is that more traditional pixel based ⁓ like object detection classifier like would be reasonably likely to say like, okay, yeah, I got E1 and I got 75 and I got like most of this shape, like yeah, it's pretty good. Seems like I found it here. what's Elan: As you can see the You don't need to mark up as many ⁓ pat ⁓ factors as you need as you did right now. You ca you can just pick a few, but the fewer you pick, the less accurate it will it will be because it's not as unique. Mike Powers: Mm-hmm. This one's smaller. Elan: So the reason it found this specific ⁓ shape, if you zoom in on it. As you can see, the ⁓ actual author he did not include the ⁓ other ⁓ vector underneath. This other vector is a part of the the the the pipe or the wall, which is not ⁓ captured in the analysis. This is Luigi La Corte: Maybe where Andrew's maybe where Andrew's going with this is so like this this looks like a this is a workflow in a way. Maybe there's like via MCP, it's here are all the items, which ones do you want? ⁓ but relative to what an estimator, a human estimator would do if they spent two weeks on this, like how how well do you perform? Because I'm sure you have your own internal benchmarks. Elan: So in terms of the geometrical models, there ⁓ as I said, it's it's a binary thing. It's ⁓ it's ⁓ it's a yes or a no. ⁓ essentially ⁓ are is there a pattern match ⁓ in between ⁓ the the the this shape and the other shape? ⁓ the the pred Luigi La Corte: Now that I that I understand, but it's kind of sidesteps the question, right? Like is if you have if if a human captures a hundred percent of these E1s, how many does Kamae capture? And like you're gonna have like it's not you've chosen to use geometric geometrical mod models or however you've done this, right? But there has to be some relative performance, irrespective of the model that you've choosed. You've chosen because that's what the customer cares about, right? They're like, how many did you find, how many did I find? Elan: ⁓ the Jamaica models are a hundred percent accurate, ⁓ or ninety-nine point nine percent in our benchmarks. ⁓ the only

things that can ⁓ make it not accurate is ⁓ inherit ⁓ geometry that can be different, that the architect ⁓ maybe placed what this e1 for this is a great example for ⁓ in that ⁓ in that perspective. You have this E1 and it's the same grill as the other E1, but it actually missing ⁓ a vector, which makes it different completely to our model to predict. ⁓ so in those cases it will not predict as well. But ⁓ otherwise it will predict and find everything else. So for example Mike Powers: And it just allows you pr the way that we're thinking about it is it's like it's the merge of the two between an LLM and this. And that's like the checkpoint, because like the LLM's gonna the LM's very very good like Opus is very, very good at counting, hey there's you know, whatever seventy seven S ones on this project. And then it's like, hey, I've got seventy seven S ones, I have the like the zone of where this is where this is projected. How do we harness that in to actually get the exact bounding box around the piece of equipment? Luigi La Corte: Mm. Mike Powers: Or around that print and those two together, 'cause like you're right. Yeah, like it's not it's like for the estimator, it's like if I just use this, it's gonna miss this E one if I'm only like doing capturing this E one up here as an example. But the merging of the two together is like very, very accurate at scale. So like and for us it's like still trust but verify, and I'll show you in the app. Luigi La Corte: I can see. Yeah, because like in in a ben in a benchmark issues will, depending on how you set up your benchmark, could be irrelevant, right? They'll effectively just say how many E1s did you find irrespective of the geometrical shape. And if there's 89, but you found 79, then you know that's your numerator denominator. but yes, you're right. Like I'm assuming because of the way you've implemented this, that would be it's still a hundred percent accurate. But from the customer's perspective, they look at it, they're looking a little bit different. Andrew Zukoski: Do you ⁓ like I feel like I'm it's been a while since I did any geometry stuff, but in olden times I had a little exposure to it. ⁓ the I one of the things that I feel is always been really interesting working with LLM technologies as a software engineer is that unlike many previous technologies, it's actually like they're they're not zero to one. Like if you do Elan: So we are Mike Powers: You Elan: Ha ha. Andrew Zukoski: like a ⁓ vector embedding and a search, or like what's closest to this? What's the thing that's most like this? You always get an answer. if you ask Claude for something, you always get an answer back. How good it is, like, I don't know, but it does mean that technology's very, very resilient to sort of slightly fuzzy things, right? That typos and so forth, like ⁓ handled by the system like very cleanly. And so I'm just like cu does that like I can imagine in in geometrical stuff there's like, okay, like the path is there, but like actually let's say it's two components and they end up being placed like slightly varying spaces apart, or like rotations or reflections, ⁓ dyspatially. And so how does ⁓ If you've got this it seems like there's like an exact vector based representation of sort of targets that are being recognized, like how does it and maybe maybe you say that's the secret sauce is like it's like you've embedded in it logic to check reflections and these other slight variations and so forth. I don't know if that's a I don't know if there's question in there. How does it work? Elan: ⁓ it's a good it's a good question. It's a good question. It's as as you said, the models and ⁓ AI generally speaking, machine learning is not a hundred percent correct. the thing is about the thing about geometric models and specifically about the the kind of data we digest, which is CAD information, it has an end. Meaning let's say you are trying you take you're taking a picture of For example, of a parrot, right? And you're trying to t teach a model, okay, this is a parrot, and this is another animal which is a dog or a cat or so on, so on. The model will always produce you an answer a certain confidence level, right? And you try to segment it ⁓ in a in a computer vision manner, if you're to let's say you have a parrot and behind the scenes you have ⁓ the the blue sky, right? If you try to treat it like and you have to treat it like a ⁓ like an image, right? Pixel based when you zoom in and you get a win. Wavy kind of representation of the pixels essentially alien and ⁓ everything else. So you don't really have a hundred percent or anything, right? It's it can it can reach like 99 points something, but not really, and in the actual production environment. In geometry, in geometrical models specifically in CAD joins, there is an end, meaning there is an actual vector the screen and nothing else. So it has to reach this kind of ⁓ geometry to the hundred percent. Once it reaches that, there is nothing behind, nothing more. As you say, ⁓ one picture can ⁓ can mean a thousand words. In geometry it's not. line can mean one line, two lines can mean two lines, and so on. or or a specific geometrical property and Andrew Zukoski: Yeah. Elan: the the constraints that it has with the rest of jum the the geometries in the drawing, it always has an end. So this is the unique thing ⁓ that allowed us to reach this kind of accuracy. Andrew Zukoski: Okay. And so how do you how to Did d I mean, I don't know, do dimensions still show up? Like the last time I was like on physical things, ⁓ y you would dimension shop drawings, right? And so there would be a drawing, a PDF, like a construction drawing or something, and like the shape was what the shape was, but written dimensions, like, were source of truth. And I mean this is one of the biggest issues, it's like one of the first health checks anyone should implement in Revit is like are people overriding the measured dimensions? Because

it happens where columns will be laid out at five feet in the model and people say like, ⁓ actually I want it to be seven and they will type seven feet in between it, such that if you measure like the drawing, like without referencing the written dimensions, you're not going to work against the intent of the drawing. Like, is that like still the dimensions still rule in the mechanical space? Elan: Mike, I think you'd you'd better a answer it. Mike Powers: Yeah, yeah. So for us it doesn't matter because we're not doing like linear takeoffs of ductwork ⁓ in the application. So like I just need to go count the but I need to go count the grill to say that there's a thousand grills on this project as opposed to one grill on this project. for and I this is where like even for like a mechanical subcontract that's doing the ductwork takeoff, they're still not gonna really rely on the drawings ⁓ for most accurate. Andrew Zukoski: Yeah, okay. Mike Powers: takeoff for that specific reason because the drawings aren't accurate. the the comment I always had back, you know, so I can reference Turner is the mechanical estimator at my office in Philadelphia would always tell me about the penetrations through slab. Engineers never put the penetrations through slab as you're jumping from floor to floor. So he would add six to eighteen inches on every floor and every takeoff, regardless of what his tool said. Because he knew that there wasn't the pipe, there wasn't the ductwork, there wasn't anything that was there. So those were like kind of like these like things that he had in his head. I remember sitting with him and doing this. ⁓ and he'd have that pen that fun pencil that you like set the the the the scale up to and you'd like literally just go, ⁓ that's a hundred yards, and then you'd write out an Excel sheet. so I think that's that challenge is still very, very hard because there's not a actual answer. because it's not actu actually a actually answered on the drawings and when you sign a contract with a subcontractor, but when a when a general contractor brings out a mechanical sub, they might they're not gonna say, Hey, I only have, you know, ten thousand linear feet of two by four ductwork. that's just not like it's not really a thing. Like you like you own drawing intent. So they're gonna put fluff in that number regardless. So hey, we're at a hundred based on our estimate. We're gonna bump it up by twenty percent. we're gonna carry that number. And there's also conflicts when you're like actually like you gotta remember, like in the HVAC world, as you're laying things, HVAC usually goes first because it's the biggest in when you're actually like installing in a field or when you're doing BIM modeling. you lay your H ⁓ HVAC because you can't run electrical conduit through an HVA duck work, then you go like to electrical, then you go to sprinkler and you go through that process. ⁓ There be changes during that as well, right? Like, so you're actually linear

feet changes because they switched the They switched the light fixture style on an RFI. The light fixture now is six inches deep in the ceiling as opposed to three inches deep in the ceiling. So now I have to actually spread my ductw this way as opposed, you know, now it's eight feet, eight inches wide as opposed to six inches wide and four inches tall. So there are all these things that happen during the coordination process. So to answer your question, like it's hard to just give, like, hey, I'm I'm buying 10,000 linear feet of two by four ductwork, because it doesn't really. Doesn't actually do anything for the customer. So again, like saying, hey, we're 100% accurate doesn't really matter. And I would guess if I was another software vendor doing like, if I was a toggle ⁓ this I would like have the functionality to add those flop numbers in there, right? So hey, we did this linear takeoff. I know that there's a turnup. I'm gonna, ⁓ there's 17 floors here, there's 17 ductworks running through floor, ⁓ 17 times times I'm gonna add six inches. So now I'm gonna add additional X in the application. Andrew Zukoski: Just and Elon, just like Mike Powers: Stuff like that. Elan: So that already comes into what you just said, it all it it comes we are reading what's written, right? The architect, the ⁓ HVAC engineer, anyone who put those documents Luigi La Corte: So Elan: We're just reading them, right? This is our part of the of the flow. If you wanna edit things up, if you want to add, as you said, ⁓ sixteen inches ⁓ towards ⁓ each slab, it's it's your own thing. We're trying to provide you actual precise information and actual measurements of what you got, and then could do whatever you want. you can call it also bid engineering. So our design partners, our ⁓ customers, they I'm speaking to them a lot. They Luigi La Corte: Okay. Elan: actual ⁓ consortium companies, not the software ones, they explaining to me like a bunch of you know secrets in the in the industry right who wins the actual bid like if everybody get the numbers right example So what's the difference? would you actually win the bid, right? Everybody has pretty much almost the same price ⁓ as from vendors, everybody has if you and you have if you have the actual measurements, so what's the difference? So the actual secret sauce in that is exactly that you try to find vulnerabilities essentially in those documents, those specs, in the bid to to leverage that and the bid. I have a great example. One of our customers, he was building he had a $50 million project. Okay? And it was a huge skyscraper in the Middle East. Essentially, the bid was everybody was bidding pretty much around this 50 million range. He figured that he can replace a bunch of steel beams with another material, and he approached the developer, told look. I'm going to place my bid $10 million less. Or I don't remember the real number. Like let's say five to ten million dollars less a lot. Yeah, it's like steel. It's like a

it's the it's a huge material. It's a structural material, which ⁓ which takes a lot of ⁓ budget from the project. going to place like a five million, ten million dollars less, but ⁓ if you let me change the design. The developer is like, yeah, sure, if it if it's up to code and it's ⁓ if it makes sense, why not? Andrew Zukoski: A lot. Yeah. Elan: So he underbid everybody by five between five to ten million and pocketed another five just because he was so knowledgeable about this topic that he could do it. That's like the real big money about construction. This is like where GCs make a lot of money or or subs that make a lot of money when they know how to re-engineer Luigi La Corte: Mm. Elan: the actual drawing. Because essentially architects and structural ⁓ firms, they are very good in theory, but ⁓ they're lacking in practicality. This is where like the big good, knowledgeable GCs come in hand and come in play. Luigi La Corte: Does it make me think that as quantities become Andrew Zukoski: Okay, so I feel like it's clicking for me. Kamai offers call like document intelligence. Like you enrich a PDF with by making information about what that PDF represents like more But there's like another layer between interpreting that information and having the design intent. Like adding the fluff working from dimensions instead of from measured, like you know, linear quantities, that sort of thing. There's like another layer where the vendor that's working with either the customer that's working with Kamai via Claude or whatever, or the vendor that's working via the API is sort of responsible for that. Is that is it starting to click together? Am I getting it right? Okay. That's helpful. Elan: Pretty pretty much, yeah. It can be either through our product, it can be either through a vendor's product, it can be through ⁓ an anyone's product, 'cause essentially it's a custom thing you do, right? Another cool thing is we are going to pretty much soon release our entire u user interface open source. It's you know, you know, in a world where cloud dominates and ⁓ ChatGPT dominate code, front-end applications are not really scarce in a way, meaning it's another wrapper around whatever we built and our core value. And that will us ⁓ to essentially distribute software. much better in ⁓ in a much better way. You can install it offline and just our API calls to do your stuff. And this is like the exact ⁓ go-to-market and and ⁓ a business model of a result-based outcome so you're you're you're paying for results essentially not for a seat ⁓ per se. Andrew Zukoski: Great. Yeah. It's like per per call. sorry, I'm like I feel like I'm playing catch-up. coming back to the example of the general contractor, contractor who proposed a design change and won the project and made a bunch of like every everybody wins in that example, right? How does Kamai help that happen? Elan: Exactly. Andrew Zukoski: Like if that's the prime profit driver, those sorts of decisions and interactions is the prime profit driver for the general contractor customer, like how does Kamai help that happen? Elan: good question. When heard this story from a GC, the guy was He he was 80 years old, still working hard, had 60 years of in the right? And when I was it it was before we had a product or anything else. when we were ⁓ me and my co-founder were questioning him and trying to pick his mind. and understand what's going on inside his head, we're like, ⁓ my god, we need to replicate him in a computer. Okay. But ⁓ of course it's much harder, ⁓ it's easier said than done. and we figured okay the deterministic thing that we can actually achieve here is to get the actual numbers because it just slows down everything else. If this guy had the exact numbers Mike Powers: Ha ha ha. Okay, so Elan: exact measurements, exact everything, and now he is only focused on how to re-engineer the project to make the most profit. That's changing that's that's like a a different kind and different way to approach everything in general, right? So the big GCs, the big boys are hiring internal structural engineers, they're hiring internal people who are redesigning those bids ⁓ because build they're bidding on billions, and if doing such a thing, a 10% increase on a billion dollar ⁓ job is massive for a council for a company, right? For such such such a project. So it eliminates the time to do the the the I would call it the dirty work, right? to actually do the art, which is the actual estimating, which is the actual engineering, which is the which is the actual creative part where you put your effort and this is how you win jobs. So this is where I position Kamai. I eliminate the deterministic stuff so you can focus on the beyond. Luigi La Corte: Yeah, it's a great val it's a great value proposition. Like everyone complains about takeoffs. Every small business spends too much time doing takeoffs for projects that they don't win. It's an absolute no brainer. For for smaller companies that are doing, let's say, public tender bids, what do you think happens then? Right? They can't submit a non compliant bid to, you know, the municipality or like the city of New York, maybe. How how do you think that shakes out? What what do you think happens there? Mike Powers: Tent tenders or bids in Amer in in translation from Canada land to to American land. Andrew Zukoski: Not the chicken ones. Luigi La Corte: So what do think what do you think happens there, Yvonne? Elan: I will give you another ⁓ I will give an example about my company from a different industry. So from a software perspective, before let's ⁓ a year ago, right? A year ago I had a software developer full-time working on on our user interface. why? Because essentially the answer later on was the models, the AI models were not as good as they are right now. Later down the road, ⁓ seven, eight months ago. We just I replaced and repositioned this developer, who is a very skilled developer, towards backhand and

more important core jobs. And I hired a lower skilled ⁓ and employee Or even myself, I also did it ⁓ a lot in front end. ⁓ actually this is I I do know how to code it for front end specifically, but essentially it made it so much easier and I just elevated a person to another position because AI could eliminate this friction point. AI essentially solved front end for software developers, right? And you utilize those tools and elevate yourself to do much higher rewarding jobs. Same way it's going to happen in construction. If takeoffs are eliminated, you're forced to do other things, right? You're forced to do other creative things, you're forced to figure out, okay, now that the takeoffs are given, what can I do to win the job? And it's beyond AI, it's beyond other things. Of course, you you will utilize AI tools and you will utilize other things, right? But you're just eliminating this annoying part that you shouldn't be even doing from the beginning. So this is this is the evolution of I would say pre construction at least ⁓ from my perspective. A a a takeoff manager could become an a an estimator. That's like that's like the the actual if you put things in terms. ⁓ Luigi La Corte: Right. Yeah, I think I think, you know, Mike has this this talk track where the CM fee goes to zero. This is not necessarily relevant because we're talking about lump sum, but margins go down. But I actually do think you're I think you're right. It'll force people to become creative. They'll either have to vertically integrate and start competing on projects where they have an advantage. Okay, so I'm gonna bring an electrical contractor in-house, I'm gonna only bid on projects that have electrical, and then I'm gonna go after the electrical subs margin on those jobs. And I'll be competitive there. And I think that's what's gonna it's gonna it's gonna force a lot of vertical integration from from G C's to major subs. I think. And then you're right, like I think VEs as well. Like everyone's gonna have to come up with really creative ways to save money and hopefully the owner's receptive to that. But I think it's ⁓ Elan: Yeah, but it essentially i it it it's already happening. Meaning you are say you're a GC, you're doing bridges, right? You're doing civil ⁓ civil works. You're not going to bid on a skyscraper and vice versa. So you are already filtering down what you can and cannot do, where you you have the unfair advantage against your competition or not. You are already doing that. It just accelerates it and make it ⁓ it makes it available to the the smaller guys, right? That's like that's the actual advantage. Luigi La Corte: Yes. Elon, are you are you Elan: If if if ⁓ i if a guy who knows ⁓ a specific ⁓ geography, for example, I'm a GC in America, in New York, in in ⁓ Long Island, I know exactly ⁓ the ⁓ environment, I know the ground, I know the the vendors, I know the routes, I know everything about this location. zero surprises right? Mike Powers: Yeah. Elan: I

have an unfair advantage of knowledge compared to another guy who is l living in Pennsylvania ⁓ bidding on jobs ⁓ from his computer using Kamai, for example, if ⁓ takeoffs are a given, ⁓ If I'm just clicking, clicking, clicking, clicking a takeoff, takeoff, bam, bam, bam, sending a bunch of bids. A guy from Long Island doing a Long Island job would outperform in any way in like ⁓ a percent of the time. That that's that's the reality. Luigi La Corte: Yeah, I can see it. Mike Powers: That just settles back to like the person thesis, right? Like it's actually individual, like the people that are the the biggest assets opposed to any like technology advantage that that a a contractor would have. Luigi La Corte: Cool. and then and then last question, ⁓ where does come where does Kamai go from here, Elon? What what do you focus on now? Elan: Our biggest focus right now is to integrate ⁓ so Mike has represented my product. So funny. doing that and not me. but just shows how ⁓ easy it is to ⁓ to use it and incorporate it. We are very heavily focused on creating essentially a very structured deterministic database about the project and to digest all of the documents into one place, which means you could query and ask questions about the whole project as one object so you can task your from mcp your lolm hey I'm doing this job and then you start doing your flow as you would do in cloud chat gpt or even through our platform or another vendor's platform. this will provide us when we are and this is our ⁓ like huge focus for us right now, a more broader a ⁓ scope of information in between pages ⁓ and in between more ⁓ features like the beyond geometry, text and all that, into one place. Once it's a one warehouse you can ⁓ ask questions about that's like that's where Kamai is headed, to be this verifiable information silo. Luigi La Corte: Yeah, that's exciting. So right so right now like the the symbol legend doesn't talk to like five pages Those those things don't communicate right now. Elan: Yeah, right now yeah. The is really per blueprint, ⁓ rather than the whole project. Luigi La Corte: Yeah. I see. Yeah, that makes sense. Once you do that, it'll you can automate a lot, obviously. You you you won't need to draw symbols. You already you already know them. That's amazing. Elon, this has been great. really excited to see where Kamai goes. I wanna I wanna have a conversation as well about how it could work. We're evaluating a few other API partners for takeoffs. We don't intend to build takeoffs, but we can keep that for offline. Thank you so much, Elon. We're gonna go offline. Mike, any ⁓ Mike or Andrew, any closing closing words or or Elon, anything you wanna say? Elan: Exactly. Awesome. Pretty exciting to do construction nowadays. I think AI and LLMs and everything else is it enabled us to advance so far. We are four founders in construction. Andrew, ⁓ you're a pretty much a veteran like you too. So you guys know that ten years ago this podcast would not be even possible, right? You would find founders in construction tech in the world maybe. And now it's ⁓ it's it's good time. Andrew Zukoski: Yeah. Mike Powers: Mm-hmm. Elan: It's it's advancing and it's awesome. Luigi La Corte: It's amazing. Okay. We'll close it there. Thanks everyone. See ya. Mike Powers: All right, I'll Andrew Zukoski: Thank you for coming on.