Latest / Construction Technology Podcast / Drawing Benchmarks, Unclouded Changes & AI-Native Workflows: Inside Trunk Tools (with Chris Boyd)
Transcript
- Luigi La Corte: Hello everyone, welcome to the construction technology podcast. I'm Luigi, CEO at Provision. I'm joined by my co-host Andrew at Join, Mike at Build Vision. We're joined by special guest, Chris Boyd from Trunk Tools, Chief Product Officer. Chris, how are you? Welcome to the show. Chris Boyd: Doing well. Glad to be here. This is gonna be exciting. Luigi La Corte: Yeah. Thanks so much for coming on. Maybe, you know, how you came on was interesting. And a couple of people messaged us and said, wow, this is gonna be a really fun show. How did this come about? Chris Boyd: Yeah. I think the the origin was a LinkedIn exchange about the just how valid are all of these frontier models and their ability to understand and comprehend construction drawings. I know there's a lot of opinions about that. So we've done a ton of research on this. We built out a whole team around this. So yeah, I'm excited to have the conversation so that folks in the field understand what's real and what's not. Luigi La Corte: Amazing. Okay. So maybe give us the overview of trunk tools and then your role at trunk tools and then we can get into the questions list. Chris Boyd: Yeah. Yeah, Trunk Tool has been around for a little over four and a half years. Ironically, started with a different product focus around incentivizing labor financially. But we incubated the product that's now our flagship today about four years ago. And the problem we set out to solve was how do we apply AI to construction in such a way as to make the data specifically more accessible for the folks in the field. So that started by building a rag, a retrieval augmented generation platform for construction, expanded into agentic automation, comprehending drawings, and expanding. on that pretty significantly. So ultimately our goal is how do we make the data more accessible so that folks in the field can focus on building rather than doing what we describe as document processing, which is trying to figure out how do these things differ from one another, how do they clash together, where are they disagreeing or agreeing, etc. So yeah, we've been really focused on the field and then we've been expanding that more and more into the back office. Luigi La Corte: Amazing. Okay. So the first question we had here was around a I think one of the LinkedIn exchanges that, you know, Mike is is fame f somewhat famous for on LinkedIn, you know, challenging ch challenging these posts and making sure that, you know, people are are honestly saying what the performance of their tools are. I think there was a post from Sarah around and Sarah's the the CEO of Trunk Tools, around the performance of your I suppose you have a a computer vision model that is better at identifying objects than you know these these let's call them like generic LLMs right like from open AI and and anthropic and whatnot. I think Mike, you would ask for a benchmark. Sarah says we only provide benchmarks under NDA. The question was well benchmarks, you know, they serve multiple purposes, right? It's a yardstick to measure performance first and foremost. But in today's day and age, it's really also a marketing tool. And so the question that comes to mind is great, if if it's so much better, why not make it available? What's the what's Chris Boyd: Yeah. Luigi La Corte: the response there? Why is it only available under NDA? Chris Boyd: Yeah, so for a little while there, the reason why I was only available under NDA is because a lot of those benchmarks were accompanying the actual drawings that we were using for the testing, right? And so all those were were customer drawings. We we only train on drawings that our customers have given us explicit permission to train on. So that was part of the reason. But but I'm bringing for this conversation some of the data that we have related to those benchmarks and a lot of the work that we've done over time. I think yeah, one of the things I know everyone's excited. I think one of the things that's really important though is like we don't have a particular allegiance to the solutions that we've developed around drawings. Our system is extremely modular, so we can hot swap models in and out. We use right now about 21 different models over the full range of our platform. And you know we are testing new models that are released on the frontier side on a regular basis. So I I have a bunch of data that I can show you both on the progression of models over time, which is really interesting in some of these specific scenarios, and then also how our system stacks up against that, specifically in the area of object detection and a few other examples. So I'm excited to dive into those. Luigi La Corte: Sweet. Yeah. Why why don't we do it? And then maybe if Chris Boyd: Let's do it. Luigi La Corte: you even take a step back, what do you define as model? Right. Is it the full end to end from ingestion to output? Are you talking about one of those 21 processes when you say our benchmark? What are we specifically Chris Boyd: Yeah. Luigi La Corte: talking about? Chris Boyd: Yeah, it's a great question. So I think model is like agent, it's an overused term in the industry. But when I'm thinking about model, I'm thinking about these discrete jobs to be done where we're leveraging either a computer vision model that we have d developed or trained, a frontier model, LLMs, VLMs, etc., these decision models that exist now, et cetera. So all of those different models apply across that entire range. And so the only reason we've ever developed models internally are because we're dissatisfied with the results of a model that's available externally. And so as we've had situations where a model we developed internally, whatever, two years ago gets replaced by something on the frontier side. And so we're more than happy to do that. Yep, absolutely. Luigi La Corte: Sure, happens all the time. Yeah, of course. Yeah, the only thing you I'm sure you care about is getting the best best result. But when you say when you say our model, you are saying Chris Boyd: Yep. Luigi La Corte: our proprietary model. Chris Boyd: Correct. So I'm I'm referring to models plural that we have developed internally for specific jobs to be done. So I know one of the topics is gonna be about object detection. So that's not that is not a singular model that's responsible for that. That's multiple models that we've developed where if we were leveraging a Frontier L L for example, to do that work, it gets only so far and that's where we have to augment with stuff that we do on our side. Luigi La Corte: Right. So maybe we can start with the headline. Because what also wasn't shared with w in Sarah's post was the number, right? So what is Chris Boyd: Yeah. Luigi La Corte: what is the number, right? Benchmarks have numbers. What's the what's the what's the trunk tools number on object detection versus the rest of the pack? Maybe you have like a you have a visual to accompany this. Amazing. Chris Boyd: Yeah, so we think I do have a visual. So yeah, I'm gonna I'm gonna share my screen and I think it'll help visualize some of this. I think one of the things that would be helpful here as well is to talk about really quickly like what do we mean by drawings in the first place and reading drawings. So I I have this visual real quick. Let me know when you can see this. Awesome. So we think about, if if a general contractor comes to us and says, Hey, do you read drawings? Our first question is like, what do you mean by reading drawings? Because there's a whole, you know, gamut here. And so we think about this in different levels of complexity, right? You start with just version control and stuff that's very simple. And then you move up to like how do you connect one drawing sheet to another one and linking those? How do you create the ability to do takeoffs? How do you leverage this for procurement and shop drawing reviews, all the way to as builds, constructability reviews, et cetera? So to be clear, we don't do all of these things yet. These are all things that we We aspire to do, but many of them we we've developed up to this point. So, to your question about what's the number, the way we think about this is at each granular level what makes sense to measure. So we'll talk for a second about object detection. So that kind of fits in this ask and measure space. so this is gonna be a little bit nerdy, but just roll with me here. I'll I'll scroll down to show you some actual values here. So here's here's an example of different objects that exist on obviously tons of construction drawings. I know this is a little bit noisy, but the number that we're looking at is the F1. So again going super tech co from it. Luigi La Corte: Wait, we're still seeing the I'm still seeing the twenty six jobs. Chris Boyd: Alright, how's that? Can you see that now? Luigi La Corte: great. Perfect. Chris Boyd: Beautiful. So if these are objects on the left hand side that exist in construction drawings, the idea here is how do we measure our ability to detect those effectively? And so some folks may be familiar with F1 scores, which is essentially a combination of precision and recall. And so what we've done here is every single one of these symbols represents a lot of different frontier models. So we've got 6.1 Terra in here, we've got Argon. We've even got Claude Fable 5.2, which just came out. We're actually running this with carbon that Gemini just released as well. And so where our kind of cortex models fit is this vertical white line over here from an F1 perspective versus all of these other models. and so when you ask about what's the met the number that we're measuring specifically for objects, it's down to the object level, the F1 score for each of those object types. Does that make sense? Luigi La Corte: I see. So we're looking at intersection of union. So basically, the how much of the box is overlapping relative to your data set. And Chris Boyd: Correct. Luigi La Corte: this data set is what presumably like hand labeled, internally QA'd. and the this white line on the end here, this is the court. So this is cortex, this is your model, this combination of of models that gets this performance. Is that right? Okay. Chris Boyd: Correct. Yep, that's correct. Yeah. So the the idea here is to and we obviously are constantly expanding this. Our target is to be at about three hundred and fifty objects that are call it for the 8020 rule that exist at construction to have F1s that are the greater than 85 to 90. And so we're constantly having to expand that data set. And obviously as you can see a lot of these are more common. So this was run against about 118 sheets specifically drawing sheets to kind of assess where we stand. And these are all quote real drawings, nothing synthetic here. This is all based off of what customers in the commercial construction space are utilizing on a daily basis. Luigi La Corte: R right. Okay, cool. So obviously these are all towards the right. So these are very good relative to the model. I guess the biggest discrepancies between general models. Yeah, why why don't you walk us through it? And maybe t tell us about how like what's the what's the impact? Like why is this why is this material? Chris Boyd: Yeah, so a and I've got so much more detail here than folks probably even want to see. I even have images at the bottom that show these practically and how they differ. but I would say what really matters here is that ultimately the folks in the field need to have something that has a high degree of accuracy for anything related to drawings. And this is the reason why we've held back, honestly, a lot of the stuff that we've been building, because we're not satisfied with the accuracy of certain things. And folks in the field should have that high bar as well. You know, we have this saying which is that folks in the field need to be able to focus on building, they shouldn't be focused on, how do I prompt this correctly to get this one little thing out of it that I need. And that's part of what we're trying to accelerate here. the other thing though that I think is really interesting, I'm going to scroll back up to the top, is this notion of cost because you know the approach that we're taking here in leveraging these exterior external models for the running these tests is that we're trying to make a pass, if you will, with each of these models each time to see how effectively it can extract this information or draw those bounding boxes around objects. And the cost goes up significantly depending on the model that you're using. And so Luigi La Corte: Is this to is this token cost? Chris Boyd: Correct, yeah, this is all related to, okay, I ingest one sheet. You can see this cost this is a cost per sheet across ultimately 38 different objects. So to run those 38 different classes of objects through each of these models, this is your cost per sheet in order to get to those numbers that you were seeing down below. Luigi La Corte: So this is just to be clear, this is the cost of the tokens multiplied by the token price, but not necessarily the cost of let's say a two hundred dollar month codex plan where this is not completely passed through, right? Cause you know there's like a you Chris Boyd: Correct. Luigi La Corte: twenty to one subsidy. Chris Boyd: Correct. Yeah. Because we have to look at this from a volume perspective a little bit differently than an individual who has one of those licenses sitting in a trailer. So, you know, we have millions of files now that are going through the system every single month. So we have to create a solution that is not only highly accurate, but also is gonna make it scalable from a cost perspective. Mike Powers: Gotta push back a little bit on this cost. So we're we do a thousand projects a day. Chris Boyd: Yep. Mike Powers: a full construction sets, ranging for a thousand page, five thousand page construction documents, three thousand page specs. Our blended cost per project for that entire stack is ten dollars on Gemini. Gemini. So I I'm curious, like that twelve dollars per page of Argon, I have never seen that high of a cost for one page. for us Chris Boyd: Yeah. Mike Powers: to do a full takeoff. On a construction set with Opus 5.5 looping until it finds all the devices, we're at three to four dollars for five pages. So I'm gonna push back a little but that's well, that's yeah, Luigi La Corte: But is that for one asset type? Is that for one asset type? Mike Powers: that's that's one, that's GRDs. For the the fable takeoff is GRDs. For the other side, it's full construction set, full full document review. we're at $10 a project all in Chris Boyd: Yep. Yeah. Mike Powers: for the whole set, not per page. Chris Boyd: Right. Yeah, and that's not to say that if we were gonna compare up here, okay, you wanna optimize for a specific job to be done where maybe you use something like you know, a a very low cost model to isolate all of the different tiles that you wanna pass and then you're using some of those models to loop over and over again. Mm-hmm. Yeah, but Mike Powers: We're not though, we're using Frontier. We're we're doing straight frontier on all of it. No no low cost models. Chris Boyd: but so we'll we'll just talk about this specific approach here. The idea here is like for the folks in the trailer who are in many cases going to go into a chat GPT interface or claude or otherwise and upload this stuff. Maybe they're asking for a very specific thing, like I want you to count the number of doors, etc. That's fine for a very discrete thing to get done. If you're trying to inventory everything in anticipation of what they're going to need, this approach basically says you're taking a prompt that is unique to each of those jobs to be done and passing that into the model for those things. So again, I I'm sure there's gonna be differences in the approaches that can be taken here, but this is just what we're comparing it to. Luigi La Corte: Hmm. And when you say that's okay, so okay. So Astro at 140, these are for 38 classes. And again, each class is an object type, right? Okay. Chris Boyd: Mm-hmm. Correct. Luigi La Corte: So you have 38 here, you have 11 below. The total pool Chris Boyd: Uh-huh. Luigi La Corte: is how many classes? Chris Boyd: So the total we're going for, and here I'll show you. I have those 27 other ones listed, further below here, so we can look at those as well. because you can see, if you're asking, where do we struggle? you can see where the F1 starts to get a little bit lower down in some of these areas. So what we're working toward, like I said, is about 350 objects that are detected. We're at about 96 or so that are in development. And then we have at this point about 37 that we're saying like our confidence interval is extremely high. And then we're constantly improving some of those other ones. Luigi La Corte: it's it's it's very bullish. it's always to the right. So what what was with the trepidation of releasing these models? If you're getting to F ones in the high nineties, Chris Boyd: Mm-hmm. Luigi La Corte: that that is objectively good. Chris Boyd: Yeah. There wasn't any trepidation in releasing the models themselves and and functioning this way. It's more about expectation management, right? how broadly do you release it? Because all it takes going back to what I was showing a minute ago about what does it mean to read drawings is if you bring on somebody who is perhaps an an MEP superintendent, right? And all they care about is all things MEP, like a lot of those MEP objects are much smaller, they're harder to detect. So their satisfaction with what we're gonna be able to do there is gonna be a little bit lower until we increase the F1 of some of those MEP objects. So that's how we Luigi La Corte: Mm. Chris Boyd: think about it. It's like we're doing really Luigi La Corte: Got it. Okay. And when you're looking at performance here, this is I'm just curious about like the what what what is the test train spit sp split like here? is this like Chris Boyd: Mm-hmm. Luigi La Corte: an unseen document? or like I I'm curious how you actually devise these because you know there's like such variability in these document types. Chris Boyd: Yeah. Yeah, our approach is that we are taking stuff yeah the models have never seen before and we're feeding those in. We've already labeled them, but we've not trained the models on those labels, right? Because the goal is that the labels serve as the ground truth that we can compare against. But yeah, it'd be leading the witness if the model had already been trained on drawings that we were using for the purposes of these benchmarks. Luigi La Corte: We have another co host here, Andrew. Andrew Zukoski: yes, this is Iris. Iris, do you wanna say hello? Chris Boyd: Ha ha ha. Andrew Zukoski: No, not gonna say hello. Luigi La Corte: Iris, what do think about the trust the the test train split approach that Trunk Tools is doing? Andrew Zukoski: I think she likes eating pie for breakfast. Chris Boyd: You're right Luigi La Corte: Ha ha ha. Chris Boyd: Hey, I'm right there with her as long as there's coffee involved Luigi La Corte: So Chris, where where does the rubber hit the road here? So this is good performance at good cost, right? What can you show Chris Boyd: Yeah. Luigi La Corte: us what actu like so when you're looking at trunk tools internally, what does this Chris Boyd: Yeah. Luigi La Corte: materialize into? What's the so what when Chris Boyd: Yeah. Luigi La Corte: we're in the platform and you know, we're able to get this like amazing result because it's so much cheaper and it's so much better. Chris Boyd: Yeah. So we think about all of these things, like object detection in and of itself is a building block towards some of the workflows that folks are going through. So for example, if if the workflow is, hey, help me count the number of doors because I'm trying to compare that to the drawing schedule, right? That's where object detection certainly is helpful. But if you want to do something like takeoffs, you have to have more than just the ability to count and identify those objects, right? It's gotta go much further than that. So we think about object detection as one of those building blocks. Then you get into things like room segmentation, et cetera. So one of the things I can show you is Practically how some of this manifests in the product. One of the workflows that you saw earlier when I was listening or listing what it means to read drawings is this notion of comparing Andrew Zukoski: Okay. Chris Boyd: one drawing to the previous revision. So the architect releases a new bulletin. The folks in the field need to go and understand, okay, what changed here. you know, the typical workflow is that they get that new drawing set, they've got to go find the previous versions of those drawings in whatever system they're using. Maybe they open them in Bluebeam or similar and they're trying to spot the knot, you know, find everything that changed at The architect didn't cloud, et cetera. So that's one thing that we've automated fully and the object detection makes that more accurate. So I can I can show you an example of what that looks like, but that would be one practical application. Yeah. Sure. Andrew Zukoski: Can I sorry, I'm slow. Can I ask a follow-up question on the benchmarks and pricing? So it sounds like the comparison in trunk tools is like, okay, we have a we have a system that we've built that calls 21 or more models across various jobs to be done, and Chris Boyd: Yep. Andrew Zukoski: it achieves these metrics and these tasks at this cost. And it seems like the comparison point is someone in a trailer walks up to one of the frontier models like first party harnesses, they walk up to Claude or something and they upload a drawing and they ask it, Do you can we see Mike Powers: Yeah. Andrew Zukoski: one of those prompts? Just like Chris Boyd: Yeah, so I'll actually show you. I did I ran this as an experiment beforehand. This was using a drawing. We'll share my screen again here. This was using a drawing using a very simple prompt. Specifically, it was one that one of our customers had utilized that they shared with us. And all right, everybody able see this. So this is an output from Chat GPT using Astro, where the user went in and basically said, Hey, I need you to identify the objects in this. level three drawing and I want you to draw bounding boxes around them based off of what they are, right? Relatively simple. And so they they sent this over to us as the output, which I thought was interesting. So you get this selector up here of like, okay, I could pick windows, I can pick doors, I can pick all this stuff. And it's really fascinating to just look at this is like a raw output. Like if it's asking it to label doors, like you can see it just misses tons of stuff. Now, are they prompting it in the best possible way and giving it lots of examples? Absolutely not. But is this a more accurate view of what a lot of folks are gonna do and then draw a conclusion about how accurate AI models are for reading drawers? Absolutely. Andrew Zukoski: Okay, so the comparison point on price therefore is like it's it's you click Fable 2 and the drop down and or Fable 5.2 and the drop down and you say go and it brings you back something at whatever cost it does it at. Is that the like is that what Trunk Tools is competing with? Like that sort of like horizontal tool? Because there are there are Chris Boyd: Mm-hmm. Andrew Zukoski: so many other Other systems out there and other products out there that have Chris Boyd: Yeah. Andrew Zukoski: some level of logic in between the neural nets that are the underlying magic and Chris Boyd: Yep. Andrew Zukoski: the user experience that are able to coordinate and loop and blah blah blah blah blah, like inje Chris Boyd: Right. Andrew Zukoski: injecting sort of like the next level of ability. Like is that the Yeah, can you it's just interesting. Like is it makes Chris Boyd: Yeah. Andrew Zukoski: sense. Like our people rock up to chat and throw everything in and ask it for something. Chris Boyd: Yeah. Andrew Zukoski: And so that's like the basic bar, but It sure seems like a crowded space with a lot of other people promising Chris Boyd: Sure. Andrew Zukoski: something above that. Chris Boyd: Well, that's why for us, the short answer to your question is no, we don't view ourselves as competing with any of the frontier models. Like they're they're in their own space. We describe it by the way as a ride the wave strategy, which is if they're investing literally hundreds of billions of dollars in those models, then our customer base needs to benefit as those models improve. It shouldn't be like you've you've you know hitched your wag into trunk tools and you're stuck with whatever we had from however many years ago, right? So that's the first thing I would say. The second thing I would say though is that the reality of how we're testing some of this stuff is folks who go in and they have a good one-shot experience with a Chad GBT, for example. I did this with my my own house drawings, it does a great job identifying every door in my house and all the floor plans. And so if that is your only exposure to it, you go, this does everything I need to do, I'll leverage it for takeoffs on a 40-story commercial high-rise in Manhattan. That's not one-to-one. And so part of that is an education responsibility. It's not about competition at that point. Does that make sense? Andrew Zukoski: Yeah, but the if the so in a way these benchmarks serve like the primary thing they educate people with is like, hey, it it's it's like saying that you should have a SAS system instead of a spreadsheet to manage something. You're like you're you're sort of competing with the category you're competing with a category of first party harnesses, right? Like under the hood you have Chris Boyd: Mm-hmm. Andrew Zukoski: some models that you may have trained or fine tuned and probably some models sourced from like other providers and so forth, like and the point of comparison is hey, come to us. It'll be the AI will be put to better work, like some of its our proprietary AI. Don't don't just use chat GPT dot com like for these domain specific workflows. Chris Boyd: Yeah. Well the the goal here is like it's a the destination, right? The destination is It's not gonna change in the industry that folks have to interact with drawings, understand them, connect them to other things, et cetera. So if our goal is to ultimately make those workflows more effective and efficient, reduce the amount of rework that's happening because people are missing things, reduce the disconnect that oftentimes exists where there's so many discrepancies between construction documents, understanding drawings is a critical means to that end. It is not inherently the goal, right? Like I don't I w our moat is not a better F1 On identifying doors than a frontier model. that's not what we're trying to accomplish. That's just a means to an end to get there, and that's why we're building out what we have. Andrew Zukoski: Okay. What is the moat then? Chris Boyd: Yeah. I think the we look at the mode from a few different angles. One is we're meeting folks where they're at from a workflow perspective. So and I I'll show this to you in the product. One of the things we hear all the time is like, okay, I've got my my drawings, which are contract documents, but I have all of these other contract documents. I have the specs, I have the submittals, I have the project schedule, et cetera. And there's no relationship that exists between them. So when we talk about things like knowledge graph, what we're talking about is actually creating explicit connections between those other contract documents in the construction project. And drawings are Oftentimes the most simplistic visual representation of that. I want to be able to go to a drawing and make it interactive now so that something like an object tag I can click on and see the related drawing schedule and the specific row or see the related submittal for that. And so we would call that visual browsing in our system, and that's one of the things we've rolled out. So the tech behind that is one of the motes because that allows our system from a rag perspective to be better in answering questions. It allows us to understand how things change over time. It allows us to assess the Impact of second and third order decisions that are being made, et cetera. So and I'm happy to walk through some examples there of what that looks like in the product. Luigi La Corte: Yeah, let's do it. That'd be great. I think making it conc concrete to the workflow is is obviously the Chris Boyd: Mm-hmm. Luigi La Corte: the most important part. Let's see it. Chris Boyd: Awesome. So I'm show you two things here that I think are are interesting. one is w the the one I was talking about a moment ago, which is the comparison of one drawing revision that was just released by an architect and how that stacks with another version that's come out. Here we go. Everybody'll see that. Luigi La Corte: Mm-hmm. Chris Boyd: So just real quick recap on workflow, right? Typically, somebody who's in the field gets an alert or otherwise that a new bulletin has dropped and they've got to go and digest all of these different drawing revisions, find the previous ones, overlay Blue Beam, all that. So our system has automated it. So that what we're looking at right here is that this is a drawing, this is this overall floor plan for the level three that's come in. We've I've automatically identified this is ASI 12, the comparison to the previous version. We've created essentially a table of contents, which I'll open here. of every single change that has occurred. The ones that were clouded and then the ones that were not clouded. And then we've leveraged our system to not only identify the change, but then write narrative to describe what changed. And what's interesting, we're seeing this is a trend in the industry industry. Gotta love architects, they're great, but oftentimes things are not clouded and it's easy to miss things. And so, you know, here's an example, like okay, I click on this, I can see that they added this built-in countertop, cabinetry unit, chairs, etc. I can turn off these layers of okay, that was added. Let me jump to the previous version of that drawing. Sure enough, it didn't exist there before. And none of that was clouded, right? And so these are examples that they have a material impact on the cost. It could have a material impact on the timeline for the project, what the trades are going to be doing, procurement, et cetera. But a lot of the stuff gets missed. So this is one of the things that the platform automates today across the entire revision set of drawings. Luigi La Corte: How does this use Cortex? Because in this case, you're starting with the old version, the new version, you know, you have Chris Boyd: Mm-hmm. Luigi La Corte: the red and blue lines where the document isn't changing, you don't get any diffs. Like Chris Boyd: Right. Luigi La Corte: clear, Cortex, I think I'm bought into the visual browsing, right? I come to a tag, jump to the tag. Maybe that takes an agent longer to do through grep, whatever, but like I can do that fast. But this isn't, this doesn't seem to use that. Chris Boyd: Yeah, this is using things Mike Powers: Yeah. Chris Boyd: like object detection and such to help the models understand what they're looking at and what has changed. So object detection is one component of what we're doing with Cortex. this is a comparison down to the pixel level as to what has changed, but also it's got a lot of logic in it for what it needs to avoid, right? We don't need to look at things like a revision cloud that was changed or removed rather in the new version because it was present on the previous version, or they took a drawing detail and they moved it on the page a couple inches to the left, right? We need to eliminate that kind of noise. So all of that is a part of what we would describe as that kind of Cortex suite of models. It's it's easier to brand Cortex as the model when a reality is the suite of models that facilitate these workflows. Luigi La Corte: Can you s can you scroll up in the drawing? I just wanna see how you're Chris Boyd: Yeah. Mm-hmm. Luigi La Corte: how you guys are drawing these objects. So if you s if you zoom out Chris Boyd: Uh-huh. Luigi La Corte: into the page, yeah, towards the top right. yeah, so if you go to that that grid line there. Built in cabinetry unit with four chairs. Model bu sorry, modified building offset with unit module. So what is it specifically saying there is the difference? Chris Boyd: So yeah, all of these different dimensions over here on the right hand side were altered. And again, none of them were clouded as a part of those changes. So Luigi La Corte: I see. Chris Boyd: mm-hmm. Yep. Mike Powers: Please. Chris Boyd: So these are things that are being caught. And by the way, all of these narratives are being written by our system. And from a probabilistic, deterministic perspective, what's interesting is that this specific functionality that's comparing what changed from one version to the next is very deterministic. The more probabilistic aspect is that we're leveraging LLMs to describe what changed, and then we're feeding, obviously, those LLMs a very discrete set of objects and descriptions and things like that to make that simpler. So, this is one kind of manifestation. And what's interesting is we have examples and anecdotes of saying, like, I'll give you one the other day. An architect was given direction by the owner of, hey, going forward on all projects, we want you to set up this whole septic system in a very specific way. And the architect went ahead and made that change on a project that was currently in flight. That got passed, our system intercepted it and was like, wait a minute, this is gonna have a huge cost impact. And what one of the things you can do directly from any one of the changes. changes that we detect here is that you can turn this straight into an email or an RFI that actually queries the whole document set to see if it's needed or not. And so in that case they were able to catch what would have been about a million dollar change order and go back to the owner and say, are you sure you want to do this so nobody wastes any time Luigi La Corte: Mm. Chris Boyd: on it. So that's one of the workflows that we're trying to accelerate. Because I, in many cases that gets passed down to the subs because the GC misses it entirely. And that's the inefficiency we're trying to eliminate. I'm gonna show another example because you mentioned the the visual browsing stuff makes sense to you. I'll Mike Powers: Mm. Chris Boyd: kinda share a version of that here. Everybody will see this. It's basically the same view, but looking at it in an interactive way. Everybody able see that? Awesome. Luigi La Corte: Mm-hmm. Chris Boyd: So you can see what we've done here. Now this stuff is lighting up. We Mike Powers: Mm-hmm. Chris Boyd: can click on now things like related sheets. And the system is identifying okay this specific sheet so that now I could jump into that directly. I can preview it here automatically. And we've essentially created this whole table on the left-hand side of all the corresponding information. And then we can do that down to object tags. So in this case, I'm clicking on a window, object tag, and I'm finding the corresponding row in the drawing schedule for Windows. And so I have all this additional detail about that window where Typically they would have had to go and find that drawing schedule, match it to the correct row, et cetera. I also have the spec section here, so I can see the corresponding spec section to jump into that as well. So this kind of gives you a sense of of where we're going. This is where object detection becomes increasingly critical. to be able to create these correlations across the document set. So on the topic of moat, right, creating a truly robust knowledge graph of all of these different discrete entities and nodes and all of that in construction is a key, key component of that from our perspective. Luigi La Corte: Mm-hmm. So the Mike Powers: Yes. Luigi La Corte: specific job to be done here is I'm browsing here, but I want to be able to go to these items very quickly. That's done, that's done e very easily here because I could just click on the item. Chris Boyd: Correct. And then the the follow-on of that is that things like an RFI comes in, an RFI response comes Andrew Zukoski: First area. Chris Boyd: in, one of the things our our system does is assess the impact of that RFI response. So we look across the project schedule, drawings, submittals, specs, etc. We say, hey, here's the potential second and third order effects of this RFI response. You're and we we you know prioritize them essentially. Here are the things that you really need to lean into and maybe double check on because oftentimes the architect may not understand the implications of of that decision. And so that's some of the stuff that this enables us to do more effectively. And then in addition to that, this becomes a stepping stone toward things like takeoffs and estimation and everything that you need to do in pre-con, et cetera. Andrew Zukoski: So what if I'm you said the P word, which is very near and dear to my heart, but Chris Boyd: Ha ha Andrew Zukoski: maybe focusing on the like trailer field use cases, because I know that's very core to the company. Chris Boyd: Yeah. Andrew Zukoski: if I'm using this, what am I not using? Like what what presumably this makes work faster and so it's not just additive work, like Chris Boyd: Yeah. Andrew Zukoski: what what is this displacing in people's day to day? Is it boobyam? Is it yeah, go ahead. Chris Boyd: Yeah. So yeah, it's gonna be multiple so we look at it a couple ways. One, yes, there's software that people are using today. Blue Beam may be a good example for the case of that of trunk review, which is the comparison of one drawing revision to another. It may be that they don't need to go into Blue Beam to do that comparison anymore, which is great. but the other thing, I mentioned the phrase earlier, document processing. It's the stuff that's not about the specific software they're using. It's the like, okay. I got this set of drawings. Now I gotta go find the old revisions and now I've got to go op download all of them and I gotta open them somewhere. It's like all of this little connective tissue that exists between things where people are frankly wasting a lot of time. That's some of the stuff that we're trying to eliminate. So I would say part of what we are are eliminating is is time stuck, stuff time stocks that are spent going to find information first and then allowing them to get this head start into the actual workflow. you are muted, Andrew. Luigi La Corte: you're muted. Andrew Zukoski: Where do they sorry, the where do they go to try to find that information? Is it Chris Boyd: Yeah. Andrew Zukoski: like it's coming over in documents so it's like is like I would imagine it's Blue Beam, would be my guess. Chris Boyd: It's it's Bluebeam, it's Procore, it's SharePoint, it's Ignite, it's Box, it's like all these different it's email, right? It's all these different places that the documents exist. I'm getting a kick out of this whole dynamic right now. This is outstanding. Yeah. Yeah. This is yeah. This is Andrew Zukoski: Yeah, it's very very professional, right? This is this is the construction technology podcast that it's fine as Mike and Luigi are just shaking their heads and you know. Chris Boyd: what's happening Mike Powers: Never. Chris Boyd: in the trailer when the subs are coming and saying, What do I need to do to build this thing? And the super is very busy with Mike Powers: So just to go later layer deep on that, because I we we just had groundbreak and AU and again, very, very very similar pitches from those teams of hey, we're creating this connect, like this this construction brain where you connect autodesk world is issues. They think issues are like the the driving force. I would say pro core is more on the document side. like your uphill battle is Chris Boyd: Yeah. Mike Powers: my again, is it to replace a pro core for a customer? Is it to replace an autodesk for a customer? 'cause a lot of those workflows, in Procore land, if you if you talk to a Procore the this the Chris Boyd: Mm-hmm. Mike Powers: product team at Procore, they'd be like, Hey, do this with Magpie, that same exact thing you just saw showed. And I'm curious, Chris Boyd: Yeah. Yeah. Mike Powers: like, you know, right now you're obviously playing nice in the pond. Is that shift like where you I know there's rumors of not playing nice in the pond, but we can move but is the obvious Chris Boyd: Ha ha. Mike Powers: the path for you guys to be successful to have to replace a pro core in a in a in a field? Chris Boyd: Yeah. We don't yeah, we don't see it that way at this stage. The way we view it is if you think about the way a lot of SaaS has been developed over the last 10, 15 years, the notion of you put information into the platform, whether you're typing it into a form or you're uploading documents, nothing happens with that when you go to bed and you wake up the next morning. It's just being stored there, right? Unless it's in a ball and cord for somebody else to actually do something with it. The system itself is not assuming any responsibility for moving the ball down the field. That's where we view our system comes in, which is it's our responsibility to help to advance these workflows without necessarily having a person responsible for every single step along the way. And so that's where it's complementary to a lot of these systems. So I'll give you a practical example. If a subcontractor uploads a submittal into Procore, I don't want to have to rely on somebody in the trailer to get the notification the submittals been uploaded, to go in and type in the right prompt or pull from a prompt library, click a button, whatever, for that submittal to now get reviewed. Rather, that should happen automatically. It should go find the corresponding spec sections that it's relevant to, any RFIs that may have an impact on it, any drawing schedules that have bearing on it, do the assessment and give you an output, and it's just waiting for you when you wake up the next morning, or in our case, five minutes after you Andrew Zukoski: Yeah. Chris Boyd: receive the submittal. So that is where we view it as very Complimentary. Andrew Zukoski: Won't Procor pitch that though? Mike Powers: So that example just in that example, just because we had Chris Boyd: Yep. Mike Powers: that liter that we with Jeff from ProQord last week, he took the opposite position of like they want their application to be reviewed by a person. Are you Chris Boyd: Mm-hmm. Mike Powers: guys taking liability and risk on by doing that? Chris Boyd: To be clear, it's still being reviewed by a person with its output. It's ultimately just providing you with a recommendation. It's saying, I've already done the analysis for you. Here is our findings. Here is a very granular breakdown of why we are giving you this output and this analysis. What do you want to do? Do you wanna click this button to pass it to the owner architect for approval? Or do you wanna click this button to create the revise and resubmit to go back to the sub? So that's still the human component. So to your question about liability, no, we're not taking on liability for that step because the customer demand is that they wanna have that point of intersection. Could we automate that if, for example, they said, hey, From a configuration perspective, we're comfortable with it going ahead and going to the owner and architect for approval if the confidence interval is greater than ninety five percent that it is good to go. Absolutely. But that's just not where the industry's at yet. Andrew Zukoski: So Mike Powers: How would you underrate that risk if you did that? Chris Boyd: Yeah, I would not underwrite that risk. An insurance company would underwrite that risk. but there's models that are emerging for that. I think it's interesting to see some of these insurance companies that are now coming up with AI specific policies. So I'm not getting into the actuarial side of things, but props to them. Andrew Zukoski: So SAS in olden times, in the before times, if I'm like Chris Boyd: Well Andrew Zukoski: a leaf node in an org, like busy doing something, like I Chris Boyd: Yeah. Andrew Zukoski: enter there's gonna be a little princess commentary in the background. Chris Boyd: Ha ha ha. Andrew Zukoski: the I go and I enter information, I wake up the next day, nothing's happened for me. I mean stuff happens for the organization, right? Like Chris Boyd: Mm-hmm. reporting. Andrew Zukoski: Everyone else in the org is able to see an updated cash flow projection the next day automatically. So maybe we're now in a world where stuff can happen on behalf of the leaf nodes in the org. Like we're out executing on a project. I don't know where your Cinderella shoes are. the Cinderella shoes And the flow of like RFI comes in, it gets connected to everything. In the f near f near future, is that RFI still coming through, whatever the project management system is? And then where do the documents like live? It Chris Boyd: Mm-hmm. Yeah. Andrew Zukoski: is it is a school day. Here, I gotta focus on my call a little bit. Do you wanna eat your brother's Chris Boyd: Ha ha ha. Andrew Zukoski: piece of breakfast pie? You you can eat his pie too. Chris Boyd: The the way we think about that is that we want to be pretty agnostic to where this stuff originates, right? sometimes it'll originate in our system, sometimes it'll originate in the PMS. Like for example, a workflow that we have related to RFIs is that if you draft an RFI in Autodesk and forma or in Procore or otherwise, our system will run automatically to check and see if that RFI is actually needed. Because roughly 20 to 30 percent of all RFIs don't need to exist because the answers already are found in the documentation. The person just didn't know. And so that's where we can augment that way. However, there's another workflow where you can go into our system directly and say, do I need an RFI for X, Y, or Z? The same thing runs. And if it is needed, we can draft the RFI and push it straight into Procore or Autodesk. So we wanna be very agnostic in that regard based on the workflow that makes the most sense for the end user. Andrew Zukoski: Okay. And so I've having I don't even know how many years people have been coming to me and saying, like, well, won't Procor just do that? So like bear with me. Like I wanna recognize that in Chris Boyd: Sure, sure. Yeah. Andrew Zukoski: the question that I'm about to ask. But there's a there's a database row somewhere that says this is the current authoritative document set from this stakeholder's point of view. There's a bunch Chris Boyd: Yep. Yep. Andrew Zukoski: of authoritative lists of things like RFIs and change orders and communications, like Trunk Tools needs all of that. It takes it, it pulls it in, it works with it, it puts it back, is what it sounds Chris Boyd: Mm-hmm. Yep. Andrew Zukoski: like. in different places. you might not care which PMS system you use, but all the PMS systems are out there saying, well, we care one about which PMS system you use, and we think you should use our like AI workflow, you know, enhancements and automations, right? And Chris Boyd: Yeah. Right. Mm-hmm. Andrew Zukoski: Some of them play nicely in the sandbox and some of them don't. So what's Chris Boyd: Yep. Yep. Mm-hmm. Andrew Zukoski: like if I'm a if I'm a CIO at a DNR top fifty firm or I don't even know if I'd be a CIO or something, but Chris Boyd: Mm. Andrew Zukoski: why would I why would I bet on a new vendor like Chris Boyd: Mm-hmm. Andrew Zukoski: when I'm being aggressively marketed by our existing partners who we've probably have a complicated relationship with Chris Boyd: Right. Well said. Andrew Zukoski: but who have delivered for our business. why would I why would I look elsewhere? They're where they're telling a compelling story. We already have all your data and we will build the best best experience on top of it because we are only building for our like universe. Or maybe they wouldn't say that, but yeah. Can Chris Boyd: Right. Andrew Zukoski: you walk me through that a bit? Chris Boyd: I think there's lots of examples of of this in the world where, if you think about Salesforce, for example, there's lots of extremely successful businesses that attach to Salesforce. Granted in their case they give up a chunk of of their revenue to Salesforce. But I think the other reality though is that to your point, you they're being marketed something. And the reality is that in construction, folks th there's a reason why folks have the approach of like we're gonna pilot something before we're gonna adopt it enterprise wide. It's a trust but verify mentality. And so you have to actually use and benefit from the software right here and right now. And so I think that's another area where we've seen differentiation. It's like there's a Andrew Zukoski: Iris. Chris Boyd: lot of promises being made for what things can do. Not as many people saying it's living up to that promise. And so I think that just creates market opportunity. Andrew Zukoski: so we're we're better. we're delivering for your users today. Your users go put it in the field, run both of them. Your users to use trunk tools will come back and say, Bro, this is better. Like yeah, okay. Chris Boyd: Yeah. Yeah. And and we're not constrained by some things, right? Like for example, as I said earlier, agnostic to where the information is coming from. So that opens the aperture quite a bit, which is powerful. I think the other reality is just from a strategic perspective, our approach is different, you know. under the hood. We talk a lot about pre-processing. Like documents go through our pre-processing pipeline automatically, which means you don't have to rely on the user to click the right prompts or select the right documents or otherwise. Like all that's happening automatically. and then the other aspect is is pricing approach. Like we don't want anyone to feel like they're being penalized for using the system more. And in the world where there's so much consumption based pricing around tokens, et cetera, you don't want the folks in the trailer asking themselves, is it worth the token cost for me to run this submittal review? that's Mike Powers: I'm pretty sure. Chris Boyd: not the way we want people even thinking about it. Andrew Zukoski: And showing to Snow White. Mike Powers: So how does your pricing work then? What is your pricing model like? Luigi La Corte: Mm. Chris Boyd: Yeah. Yeah. So the way we think about it is it's a flat flat cost based off of the the project or projects that a customer is bringing onto the system. They can bring on as many users as they want, they can bring in as many documents as they want, they can run as many agents as they want and query it, whatever else. Like it's it's it's an all you can eat is what we would describe it as. Because we've seen the worst thing you can do in construction is create variability for a PM, for example, who's trying to watch margins and otherwise they don't want to also have to sweat. man, do I actually want my team using this fully? Because that might eat in my margin if they're using it too much. Luigi La Corte: Yeah. Andrew Zukoski: Well, presumably Mike Powers: So wait when you Andrew Zukoski: if they use it too much it's delivering enough value to make like more than pay for that. But I guess maybe not everyone people would worry about that all the same. Yeah. Chris Boyd: Right. Right. Well, and also you see with some providers that there's a multiplier on token consumption. So for example, you may consume I'll keep the a very simple example, you could consume ten tokens, you know, asking a question. But if you run an agent to do the same thing, it's consuming five to ten X as many because of everything that's happening behind the scenes. So we've seen many examples where that can run off the rails where somebody's like, yeah, I've got plenty of credits, I'm good. And then thirty days in, they're like, and they're all gone. And I thought that was gonna cover me for six months. Andrew Zukoski: Yeah, I've noticed that just with the horizontal frontier models. You know, they're my my favorite is when Claude, you know, spends extra tokens to eliminate the open AI API from, you Chris Boyd: Yeah. Andrew Zukoski: know, our application. Chris Boyd: Right, right. Mike Powers: You got you gotta stay on the the personal account. That's why stay on the twenty dollar bucket there, Andrew. Don't go into API land. Chris Boyd: Yeah. man. Luigi La Corte: Chris, I I wonder sorry, Mike, go ahead and then I'll go. Mike Powers: Yeah, I was I yeah, again, I'm just I've I've from the go I did a lot of sales, so I'm very curious how you think through this. one of the one of the pain points we hear from ProCore Audit S customers is price keeps going up. They're like Chris Boyd: Yeah. Mike Powers: owners are getting smart on the fact that, hey, I'm paying for a pro core subscription. Why would my GC Chris Boyd: Yeah. Mike Powers: and my sub paying for my pro core subscription? when you pitch a customer, are Chris Boyd: Yeah. Mike Powers: you guys seeing the ability to without saying, Hey, we're like saving you guys cost, like seeing Chris Boyd: Yeah. Mike Powers: true ROI where Hey, I'm a super, I actually need this on a job site. And I've been having, act out outcome-based returns for my customers. and how does this because like I know that the dollar for IT and construction is getting squeezed so much. How are you guys kind of like leaning front forward into that? Because I'm sure your tool what's what's the ACV for your tool? Is it like 10, 15, 20 grand a project? Like, how do you guys Chris Boyd: Yeah. Mike Powers: think about that? Chris Boyd: Yeah. So it's gonna be a a wide range based off of the the size of the project and the size of the portfolio and things like that. But what I will say from an ROI perspective is is yes, and and frankly we skinned our knees on this a little bit where we were definitely doing it the wrong way. I think when you just you just propose an ROI associated with time savings, that usually doesn't impact the bottom line for a G C because they're still gonna staff it with the same number of FTEs. It's not like they're gonna take somebody off the job, you know, that's in flight that may impact the next job that they are bidding on, but not the one that's currently in flight. So We've we've moved away from time savings and we treat that more like gravy, if you will. It's like, hey, here's the value prop. by the way, we also saves you time. But some of the things that we demonstrate for ROI are going to things like cycle time associated with the submittal process or things that are caught that were not clouded by the architects or dollar amounts from a change event perspective that were captured as a result of either drawing revisions or RFI responses. Right. And the list goes on. It's like ultimately we're trying to connect this back to operational metrics. That they should be monitoring. Unfortunately, because of PMS solutions, they oftentimes have a lot of the baseline data already. So it gives us a point of comparison. Now, does that tell them, like, my margins go improve by 2% on these jobs? No, not yet. But you can tell a story that says if I reduce my, I'll pick on submittals again, if I reduce my submittal cycle time by an average of 50% and I tie that to the critical path items that are now going to be done on time or start on time rather than eating away at float. Now you're talking. About okay, general conditions, liquidated damages that are not going to be incurred near the end. So that's part of how we tell that story. Luigi La Corte: Chris, I'm curious, from a market perspective, who do you typically go head to head against? Chris Boyd: Yeah, we have this saying in the business, which is that we oftentimes compete against market perception more than market reality, more than product realities, if you will. And so our competition is oftentimes, hey, we had a guy, right, who he was Luigi La Corte: Mm-hmm. Chris Boyd: able to vibe code this thing and it seems like it's fine, right? And so we have to talk them through that, walk them through the differences, our approach, et cetera. So there's that camp. Then there are the groups like we were talking about earlier, they're already they're leveraging Andrew Zukoski: yeah, type watch. Chris Boyd: a one of the incumbent PMS solutions. They're being told that it can do the same thing. Andrew Zukoski: Let's go watch. Chris Boyd: And so they they don't have a way to objectively measure that. So sometimes they have to pilot it side by side. So there's that reality. Andrew Zukoski: god Chris Boyd: And then the the the biggest camp, I would say, is the wait and see camp, right? It's the groups of general contractors who are like, I want to wait and see Andrew Zukoski: This one. This one. Chris Boyd: how the dust settles on some of this until I make a decision. And part of what we push for there is like I think we're past the early adopter curve, and every GC needs to be assuming some degree of responsibility for helping their Andrew Zukoski: Is that do you think we're changing? Chris Boyd: team adopt AI solutions. And even if it's not Mike Powers: Mm. Chris Boyd: And so that's where we have to push them a little bit there. So those the three camps, if you will, of competitive buckets, if you will, that we're we're dealing with Luigi La Corte: So you you never go head to head with another solutions provider, we're we're evaluating trunk tools and someone else. That that never happens. Chris Boyd: No, I'm saying that's in that second category. That's where they would say, Hey, we're looking at perhaps an incumbent PMS solution that offers a competing or what they view as competing, and we want to run these head to head and have an output. And it was funny we actually had this happen very recently where unbeknownst to us, they were running us head to head Andrew Zukoski: Yeah. Chris Boyd: and they sent us the report afterwards with the findings, which has been some of the greatest marketing material we could have ever gotten, and we didn't ask for it. So yeah, that happens all the time. Andrew Zukoski: I have one more question. sorry for the sorry for the background noise. so and sorry, I think slowly 'cause I'm on the Pacific coast and so it's really early for me here. the Chris Boyd: Well, you're s you're totally fine. Andrew Zukoski: trunk tools has raised a fair amount of money. Chris Boyd: Yeah. Andrew Zukoski: a lot of money. Big part of the cap table is looking for a venture outcome. That points to path to hundreds and hundreds of millions of dollars in revenue. Or billions of dollars in revenue as being necessary to realize that. So there's gonna be like a lot of pressure might not be the right word on board, but like that's where the that's and I get it. Everyone on this call has raised venture, right? There's nothing there's nothing wrong with it. the trick vision right now seems to be and I'm paraphrasing I might get it wrong, you know, connecting people to information to better understand the information to let the project move faster. That you're able to sort through while it documents on the project, connected to various action items that need to happen and automate many of the steps like in that analysis and follow up. Is that like in target? Okay. And so then the bet is that there's enough value and revenue potential there that you don't have to displace any of the existing Chris Boyd: Mm-hmm. Andrew Zukoski: categories of spend. That if we if we fast forward and were And Turner construction is like I have an X million dollar a year Pro Core bill. Fast forward a few years they're gonna have whatever I don't know what it is. Call it ten million in Procore spend and they're gonna have ten million in trunk tool spend. Like on top of that right next to it. Is that the Chris Boyd: Sure. Yeah. Andrew Zukoski: Okay, and that that's Chris Boyd: I it's well Andrew Zukoski: go ahead. Chris Boyd: I I'm I'm saying yes as I'm I'm just acknowledging what you're saying. My my response Andrew Zukoski: okay. Chris Boyd: to that is that I believe that the expectations in the industry, not just in our vertical, but in others as well, is going to shift around AI native applications are gonna become the default, right? The expectation should be the way I interacted with my information, my data is going to change pretty significantly. one of the things that we talk a lot about is this notion of ambient intelligence, which is in a world where you've got hundreds of agents doing something. It's it becomes unsustainable to have to go and manage every single one of those individually. So you need the orchestrator, you need things that sit above that. And so our vision of the future for construction is that there are lots of workflows that are being heavily automated that are reaching out to people when they need to get involved in certain steps along the way. That the data entry for how folks are getting information into various systems is much more intuitive. I I like this conversation I had with a a CPO of another company talking about what it means to be AI native, and his description is like, We're gonna go back to cave drawings. It's like the simple way that people communicate going all the way back to the beginning is just visually how do we collaborate on a whiteboard, et cetera, and not feel obligated to go back to the trailer and re-enter all this information into various systems. So that's the future that we're fighting for. The composition of software that's gonna make that a reality, I'm somewhat agnostic to as long as we're achieving that reality. And so we we'll continue to play nice and plug into a lot of these systems. Andrew Zukoski: Well, but trunk tools investors are extremely not agnostic to the composition of the technology Chris Boyd: Ha ha ha. Andrew Zukoski: stack that's doing that. we can play nicely with others, but Chris Boyd: Yeah. Andrew Zukoski: the you you need a you need a thick slice of that, right? Chris Boyd: Yeah, but I think a th th the reality though is that if you look at the con the spend in construction on software is a fraction of the spend on labor and labor is being heavily constrained. And so if you can earn the right to achieve some degree of that technology and and software spend, you will then, with agents and otherwise, earn the right to leverage some of that labor spend. And that's part of our vision for this is that if in a in a world where we are constrained by labor, if you need more labor. And agents can fill in some of those gaps, construction companies can expand without having to need to hire as many people. Andrew Zukoski: So if I'm Gilbain, like, right now I'm paying whatever and it goes in my technology budget. But if you fast forward you're gonna have conversations with them saying like, Hey, look, on your next project you only need to staff one P E instead of two and so you should you should give us an extra zero on our A C B per project. Is that the Chris Boyd: Yeah. Andrew Zukoski: that's the path? Chris Boyd: Yeah, we're seeing dynamics already where GCs who are doing really, really well from a earning the right to achieve new projects are having to turn things down because they can't hire fast enough to staff those projects. And so how do we enable them in the future to say, okay, for this three hundred million dollar job where you needed a PM and a few PEs and an APM and whatever, that you can take one or two off of that and still get as much productivity? So it's very much a proactive view of the future where those teams can be much more effective. We're already seeing some GCs do this with offshore resources and things like that, but I think that AI is a real unlock. It's a beautiful intersection, by the way, that we've got this massive influx of spend in all things AI, while at the same time we're being so resource constrained from a labor perspective. I think those go hand in hand really nicely over time. Mike Powers: For like for the data center build out, like the resource constraint isn't that, hey, I don't have enough PE PEs sitting in a room reviewing submittals. It's how do I have the expertise in the field of like a superintendent actually following a mechanical subcontractor? So Chris Boyd: Yeah. Mike Powers: how do you think about like those people? Like, if you go to if you go to like a large general contractor, they're like, hey, I can't Chris Boyd: Yeah. Mike Powers: find people out of college to review documents, read documents, do X, Y, and Z. It's I can't find an experienced project executive that's run a data center. Those are two Chris Boyd: Yep. Mike Powers: completely different things. Chris Boyd: Yep. Mike Powers: and how do you reconcile those pieces? Cause like my first job in college, I read the spec and I read the drawings like at Turner. Like that was the bet because then I would understand the project by layering on AI that like understands the documents. I think there's like gonna Chris Boyd: Mm-hmm. Mike Powers: be huge heartburn in the industry because then that P doesn't read the documents. Chris Boyd: Right. Mike Powers: and I'm curious, like like that, like that pitch to a Gilbane, like it's not like they're Chris Boyd: Sure. Mike Powers: having a hard time finding people out college. It's it's finding the person that has data center experience to Chris Boyd: Right. Mike Powers: run a project because Anthropic will pay you seven hundred thousand dollars to go work on a project and Gilvane can't pay seven hundred thousand dollars a year. Chris Boyd: Yeah. Yeah. I th when we look at each of these different roles on a construction project, we think about it as what up what percentage of their work can AI automate or assist in such a way to give them additional capacity back? And in some cases, you know, a PX, there's there's only ever one typically on a project unless it's absolutely massive. So that's that's not the the persona, if you will, that I would say we're gonna aid significantly in reducing the labor spend there. It's gonna be for a lot of the other individuals that are involved. And so yeah, I I agree with you there, but I also think that there's a knowledge transfer problem that needs to be solved in construction, and that's where a lot of this can help as well. Not just data centers, but how does a Gilbane or a Suffolk or whoever, how do they build? And how do we make sure that the person who is fresh out of college is equipped with that context, not because they sat through three months of training courses, but because the system is aware of that context and bringing it to bear on the project during construction. Mike Powers: There's funny, there's a premium. So Turner does Turner Camp. there's actually Chris Boyd: Yeah. Mike Powers: a premium for employees to leave after two years because they can get a ton of money because they've gone through Turner Camp, which is an interesting Chris Boyd: Yeah. Mike Powers: perspective. You like jump up a level. I remember I made Chris Boyd: Yeah. Mike Powers: my first job, I made fifty two thousand dollars a year, my first job out of college. another general the other general contractor in Chris Boyd: Yeah. Andrew Zukoski: More than me. Mike Powers: the Philly market reached out and said they pay me sixty two thousand dollars a year, just because I had the Turner name on my resume, which is Chris Boyd: Yep. Yep. Mike Powers: interesting. Chris Boyd: Yeah, we we talked about GC the other day and I asked him, how are you doing on hiring? He's we do fine. I said, Well, talk me through your process. He said, Well, typically we hire folks and we put them on job sites for a year to essentially intern with one of our people that's very well versed in the way that we do things. And then at the end of that year, they can do the same job on another job site. I was like, So you have a hiring problem. Like it takes you twelve months to get somebody in seat for some of these roles. He's Yeah, well, that's fair. So I think th there's a little bit of a frog in boiling water aspect in construction as well of people thinking that it's just so Andrew Zukoski: Della, how does I kill lesson? Chris Boyd: easy to go and hire. It's hi it's easy to get somebody in seat, it's hard to get somebody in seat with the context that's necessary to be successful. Andrew Zukoski: Okay, I have one sorry, it takes me you wanna wrap, but sorry, it takes me so long to think through Luigi La Corte: Go ahead. Andrew Zukoski: things. so earning right for technology spend or delivering on value for technology spend earns a right to labor spend. technology spend comes out of the tech org most of the time, unless it's sold at a project level, in which case, who the heck knows? Could come from all sorts of different orgs. labor spend comes from ops. Right. So the idea is we sell through technology and we deliver enough value that now technology leaders can go and say, Hey, we c inside of pick a company, we don't need to harp on Gilbain for too long. Inside of Gilbain, now Chris Boyd: Yeah. Andrew Zukoski: with this AI, I can solve the labor problem. Chris Boyd: Mm-hmm. Andrew Zukoski: Like inside the org and so our budget should go up by a fac by a multiple. my gosh. Chris Boyd: Right. Andrew Zukoski: And it's just chaos here. It's like a little construction site. yeah. Chris Boyd: The book is caught. Yeah, yeah. Luigi La Corte: I'm impressed even able to get these Mike Powers: hey Luigi La Corte: questions out. Mike Powers: Andrew, I just before you finish, I gotta hop real quick. Sorry, I got pinged on something. Chris, thank you. I'll talk to you soon. Andrew Zukoski: All right, sorry Mike, bye. Chris Boyd: Yeah. Thanks, Mike. Luigi La Corte: Okay, Mike. See ya. Chris Boyd: Good to see ya. Andrew Zukoski: Yeah. so you go to you go to IT and you say hey, like now you can ask to like multiply your budget, or is it you go to ops and you say like hey, you should instead of hiring people, you should you should sponsor a much larger like trunk tool spend. Like okay. So it's like expansion inside the org. Yeah, okay. Chris Boyd: Yeah. Yeah, it's a it's a both and. Mm-hmm. Yeah. Yeah. every company I've ever worked for, especially in construction or construction lending, et cetera, the goal is always how do you create software that makes it easier for your customers to scale and not require as much headcount to do it? And if you can accomplish that, right, then everybody's winning as a part of that process because all of these companies want to grow more. Like my goal is not to take jobs from anybody in construction. we we need more people in construction more than ever. it's been that way for quite some time, especially since the 2008 downturn. So creating more of those opportunities, making those people more effective, and then ultimately improving morale as well. if these folks are on the job site working 16 hour days and then, by the way, when they get home, they've still got to do their dailies and things like that. Like we're not actually creating an environment that is going to be sustainable. And that's part of Luigi La Corte: Great. I I love the deep dive. Yeah, Andrew Zukoski: Fascinating, I really appreciate it. Yeah. Chris Boyd: Yeah. Luigi La Corte: us, yeah, me too. one last question before we do rapid fire. What what does winning Chris Boyd: Yeah. Luigi La Corte: look like for you and trunk tools? Chris Boyd: Winning looks like the the way that folks are building changes by nature of the fact that the field teams are able to focus on building, they're spending more time out in the field, on the job site, etc. The workflows are being automated sufficiently, quick just like rapid-fire data. Documentation around construction has gone up by about 60% in the last 10-15 years. the duration it takes to build things has gone up by 30 to 40 percent in the last 15 to 20 years. And then we have this major workforce group that's gonna be retiring over the next five to ten years. And so I think all of those things intersecting means that we have to create not only more efficiencies on the job site, but we have to improve morale, we have to change the way that people interact with their construction data. For us, if we can help play a major role in making that happen, that's winning. Luigi La Corte: Beautiful. Love that. I I don't know who I don't I definitely wouldn't disagree with that. I wonder who would disagree with that. But that sounds like a great aspiration. do you feel like you Chris Boyd: Yeah. Luigi La Corte: are winning right now? Chris Boyd: I feel like we have set the stage for all of that to be possible in the future. I think we're making real impact in a positive way on these field teams. And and by the way, to Mike's comment earlier about, if if AI is just automating everything, the PE's not gonna be as good at consuming things. We've actually seen the inverse in a lot of ways because now we have folks in the field who are learning from the system what they should be doing and how they should be reviewing some things. So that gets me super excited about the progress that we're making in that regard. I think the frustrating part at times can be the what we would describe as the inertia, if you will, of the industry of, prove it to me, prove it to me, prove it to me. if the field derives a bunch of value, that doesn't necessarily mean that the executive stakeholders cutting the checks see that value immediately. Luigi La Corte: Mm. Okay, let's do a rapid fire and then wrap because I know we're past the the hour here. Chris Boyd: Yeah. Luigi La Corte: what percentage of your team salaries do you expect to spend on AI in the next year? Chris Boyd: man, I Luigi La Corte: This is internal within Trunk Tools, obviously. Chris Boyd: right, right. Percentage of team salaries. I would say somewhere in the ten percent range. it for example, we made some pretty significant changes in our go to market org recently where we have heavily automated a lot of stuff that we're doing on the go to market side and our spend went up there a decent amount, but it was still a fraction of the labor aspect. So I think that's gonna keep increasing over time. Luigi La Corte: if you could become CEO of one G C tomorrow, which G C would it be? Chris Boyd: Mm man, I would want to I won't name any names, but I would pick a G C that's in the kind of high end of the S and B or low end of the mid market where you've still got a lot of flexibility and agility and you can adopt and adapt very quickly to a lot of the emerging technology stuff. I think there are there are folks in that space right now who could be in the ENR four hundred in pretty short order if they just go whole hog into this stuff and adopting it effectively. So I think that's a really unique opportunity. Luigi La Corte: you have to choose one long and one short, anthropic, open AI, and SpaceX. Chris Boyd: Long and one short. man. I've I'm a big anthropic fan, so I would definitely go long there for a lot of different reasons. So I'll be very curious to see how this whole IPO goes. The short the only reason I would say open AI is just because I think that they're very focused on the consumer side of things and at times lose sight of some of the other areas. But again, that's just my kind of 30,000 foot view. SpaceX is its own animal. Luigi La Corte: Yeah, you can't you can't make money shorting Elon's companies these days. It's just way too hard. Okay, I think that's it, Chris. This is awesome. really appreciate you Chris Boyd: Yeah. Yep. Awesome. Yeah. Luigi La Corte: opening up the the curtain and and giving us some of the numbers and excited to see what Trunk Tools does next. Chris Boyd: Awesome. Thank you gentlemen. Happy to do this again sometime. Take care. Andrew Zukoski: Thanks Chris. Take care. Tee bye. Luigi La Corte: All right. Thanks, Chris. Take care. See ya. Bye.