Pod Insights a16z Podcast a16z Podcast | Construction Under Tech -- The Build About this episode with Martin Fischer (@fischermartin), Saurabh Ladha (@ladhasaurabh), Chris Rippingham, and Hanne Tidnam (@omnivorousread)
Continuing our series on how tech is changing construction -- one of the industries most resistant to change (and facing declining productivity) -- this episode of the a16z Podcast looks at what happens when you go from planning to actually putting boots on the ground. How can tech translate rich data sets into the just-right types, amounts, and levels of information for each different piece of the incredibly complex, dynamic, time-and-space problem that is a building site?
Transcript Speaker names marked ? are inferred from speaking time and the show’s known hosts — they are not voice-verified.
0:00 Speaker 00 The content here is for informational purposes only, should not be taken as legal, business, tax, or investment advice, or be used to evaluate any investment or security, and is not directed at any investors or potential investors in any A16Z fund. For more details, please see A16Z.com/disclosures.
0:19 Hanne Tidnam? Hi, and welcome to the A16Z Podcast. I'm Hannah, and in this episode, we continue our series on how tech is changing construction, looking at what happens when you go from the planning phase to the building phase. The conversation includes Martin Fisher, professor of civil and environmental engineering at Stanford, Sarab Ladhah, CEO of Doxel, a company that uses artificial intelligence to measure progress and inspect quality on construction projects in real time, and Christopher Rippingham, who works on technology and innovation leadership for DPR Construction. We've talked a little bit about the lay of the land already in the construction industry, some of the big challenges and new technology around the flow of information. But what happens when you get to moving into building the buildings?
1:04 Hanne Tidnam? When you get into execution on the ground and productivity and performance management?
1:09 Speaker 02 One of the biggest challenges we face right off the bat is trying to take the rich data sets that are already inefficient from pre-construction and migrate that into something that somebody on a construction job site can use or read or fabricate from.
1:20 Hanne Tidnam? And what kind of data sets do you mean by that?
1:23 Speaker 02 Like the 3D models, building information models, that kind of information. How do we make that information accessible into the field? All the way down to RFIs, written information, to plans and specs from the architect and engineers. All that information is either dumbed down usually or put in a format that is not quite as rich or usable, or has the dynamic state that, say, a data set would have within pre-construction.
1:43 Hanne Tidnam? And when it's dumbed down, is stuff stripped out of it?
1:46 Speaker 02 Sometimes, yes. It depends on what kind of information's in there. What we try and do is only give the information that's needed to do that specific task. The goal is to give location awareness, role awareness, that kind of information to a specific worker on the job site.
1:59 Hanne Tidnam? So you're kind of like fuzzing out the irrelevant stuff and crystallizing the relevant stuff?
2:03 Speaker 02 Right now, we're basically giving an entire set of plans over to a construction worker in the field. And that could be thousands of pages of information that maybe only a handful, 100 pages, are useful for that one specific person.
2:14 Hanne Tidnam? So an electrician goes into a site?
2:16 Speaker 02 Yep.
2:16 Hanne Tidnam? And he has literally the entire?
2:18 Speaker 02 Set of documents. Yeah. And so the different useful technologies now are helping Make that dataset smaller and smaller. But meanwhile, the job site is a very dynamic environment.
2:27 Speaker 03 It's really frankly crazy if you think about it. You have a group of architects and engineers that create drawings or even 3D model information just in case, just so they can sort of show we did a lot. And then the reality, they have to go through there and filter out the information that is relevant to them. And then any uncertainty in the information that exists from the design phase gets multiplied in construction. It doesn't get simpler, right? Because then you have to figure out how to allocate resources, what sequence you were going to work, you know, the site logistics. Imagine that now for 20, 30 trades coming together on any given day. And you know, that's supposed to work like clockwork, but there's uncertainties, vagaries in everybody's variability and everybody's schedule there. And then you compound that with uncertain design information.
3:12 Speaker 03 And that's why we end up with this kind of unproductive chaos that we often see.
3:16 Hanne Tidnam? What is the human result of that? Are there people sitting around waiting for deliveries? Are there people not coming up to their jobs and they're not ready for them? Like, how does that actually look on a human scale?
3:26 Speaker 02 I mean, there's a whole concept of onboarding to a job site. You know, it's hard for workers just to come out and start working that, you know, day one.
3:32 Hanne Tidnam? That's true. It's like starting in a whole new environment.
3:34 Speaker 02 They come out and they're off to the races, but that's not really the case. They really have to dig into the information, find out what they need to look for, make sure there's not other details that have been created somewhere else. Make sure that the details that they usually use are actually incorporated into the set of documents. And there's this issue of trust on a construction job site. Because like an engineer doesn't always trust that the subcontractors know what they're doing. And so they put in all these details that they think are the best details that work, but they've never installed anything in the field. Meanwhile, the field worker has to come through and look at those details and say, I can't actually build it this way. I have to do it this way. Which means we have to put in a request for information and an RFI in construction terms. And then that has to have more communication back and forth with the architect or the engineer, which adds more and more complexity onto this already complex work environment.
4:15 Hanne Tidnam? So this trust situation, that's really interesting. How variable is that? Are there different sites where there's a higher degree of trust? Or is there a general point of low trust across the board?
4:26 Speaker 02 There's definitely different contracting types that help with that. But the traditional design, bid, build atmosphere doesn't lend itself to that trust. But when you get towards integrated project delivery or design, build type contracts, there's a lot more trust that has to be established with the design team. And there's a lot more opportunity for subcontractors to be involved in the process earlier, which allows them to provide those details that are constructable and that they'd actually use in their field.
4:48 Speaker 03 This has been something that has been a bit of a surprise to me because I would say I have seen more initial adoption of advanced technologies in projects with little trust. Somebody says, oh, this is a really good tool. I can defend myself if necessary. I can be on top of the food chain because I have all the information that the others don't have, et cetera.
5:07 Hanne Tidnam? That's really surprising because you would think that-
5:09 Speaker 03 Yeah, you would think that on collaborative projects, they would say, we can do all of this great stuff. But it's not what I have experienced on projects. Then from there, it does grow into people then realize, oh, wait a minute, we could actually work together.
5:20 Hanne Tidnam? Well, let me ask maybe a kind of dumb question, but isn't the job of the general contractor to kind of be like the central, the heart of that trust?
5:29 Speaker 04 So it's a lot easier said than done. And if you actually zoom out and take a look at the construction industry, 98% of construction projects above a billion dollars run on average 80% over budget.
5:41 Hanne Tidnam? That's so much over.
5:43 Speaker 04 I know, right? Don't you just feel the dollars leaving? Right? It just feels so sad. But the reality is that general contractors get a lot of flack for that. And understanding how challenging the job is physically with human limitations is important. There's a very popular notion that, hey, manufacturing has doubled in productivity since 95 and construction has stayed flat. In manufacturing, they're doing the same thing again and again, and they've got real time feedback. So if something doesn't go as planned, you can intervene in three minutes and fix the problem. Whereas in construction, you're building this massive thing once with different people and materials every time, and you're expected to get it right.
6:22 Hanne Tidnam? And every step is different, right?
6:23 Speaker 04 Every step is different, and you have to do it all without any real time feedback. If something is installed half an inch off, it could cause a massive ripple effect on the dependencies. So there are just some human limitations in assessment of quality and progress just because of the amount of stuff that's going on on a construction site.
6:39 Hanne Tidnam? Well, let me ask you about when you say without any feedback. I mean, isn't the contractor standing there assessing, keeping track of things even with a clipboard? There's an attempt at feedback, wouldn't you say? Or no?
6:51 Speaker 02 There is an attempt at feedback. Yes, we do have quality assurance programs that are put in place and we do have people doing visual checks. And we do have a lot of walks with the architect. We have things called rolling completion lists that are things that we need to complete that the architect would like to see happen. But there's a lot of human interaction there that doesn't always capture the true intent of the design. We have the saying of measure twice, cut once. And then it's like, do you quality check that twice after that?
7:15 Hanne Tidnam? Right, right. And it's eyeballs and voices right now, basically.
7:19 Speaker 02 We do a lot of model-based layout and trying to double check that the models meet that, but it is still a fairly manual process.
7:25 Speaker 03 So in the study of 19 projects from 19 different companies who just span the globe literally. And our research has shown that people don't have feedback loops, period.
7:36 Hanne Tidnam? Wow.
7:36 Speaker 03 I mean, just informal.
7:37 Speaker 02 In all shapes, ways and forms.
7:39 Speaker 03 And certainly not on a project, not from week to week, not on projects, not across projects, not across the company, not one. Whether it was on schedule, whether it was on performance, on quality, we could not find it. There was a lot of use of building information modeling, there was a lot of using lean construction management methods, but nobody actually really formally closed the loop, to really close the loop. You've got to have a plan that matches how you're going to do the work. And that is actually much harder to do because of the scale and the one of nature. And that's where really the technology we're seeing now emerging will make it easier to create a plan that matches how you can collect data in the field.
8:20 Hanne Tidnam? So you're saying that that initial fracture kind of happens at the very beginning because of plan versus execution? And then that ripples out in a million ways and makes it impossible to get feedback?
8:31 Speaker 02 Yeah. You know, DPR tries very hard to create a plan from the initial set of documents or at the very beginning of the project, but then to continuously execute on that plan has to be revised because the job site is so dynamic and there are changes coming down.
8:42 Hanne Tidnam? Like what kind of changes? How often?
8:43 Speaker 02 So if you were building a hospital or something, you might have a job walk with one of the doctors the day before you're supposed to be installing overhead equipment and he says, no, I don't want that boom here, I want it here. And... That boom actually affects all the structural steel that's above. It affects all the MEP that's above. And so there's a lot of that dynamic changing.
8:58 Hanne Tidnam? Wow, that sounds really miserable.
9:00 Speaker 02 That's what all the general contractors to really try and manage those design changes, really work with the owner to make the building what they really want it to be within the timeframes that they have, within the budget they want and with the quality that they want.
9:10 Hanne Tidnam? So what are some of the kinds of technologies that we're starting to see come in to start knitting that initial, either that initial fragment or that are helping in that feedback loop?
9:19 Speaker 03 The big difference really in the last two years has been the ability to sense what's going on on a construction site. That's something we have never had in the history of construction, right? 5,000 years of building, things by humans, from visual to Fitbit type things to vibration to you name it.
9:37 Speaker 04 That's basically the core of what Doxel does. And Doxel is essentially a computer vision solution that uses autonomous robots to scan both the indoor and the outdoor of a construction project on a daily basis. And then we use deep learning to automatically extract how much work has actually been done and how much of that work has been done correctly. So what you start getting is this dashboard, which tells someone in finance, how many dollars of work has been installed today versus how much money are people asking me to pay? People can compare schedules, right? How are we doing in actual versus plan? And then, of course, you get the quality piece, which is, is stuff installed correctly?
10:18 Speaker 04 So it's that real time feedback system of manufacturing brought to construction.
10:23 Speaker 03 We cannot possibly maintain manually a plan at that level of detail every day.
10:31 Hanne Tidnam? So you literally tried this kind of manually?
10:33 Speaker 03 We have tried this manually and we realized that's not really possible.
10:37 Hanne Tidnam? Because you drown in the data?
10:38 Speaker 03 Well, because you've got to wait till like late in the day to learn what happened that day. And then you've got to create a plan for the next day in like, you know, seconds quickly while the managers of the site are still there. So they can look at the plan for tomorrow and say, yeah, this is good or this is bad. Because otherwise, they're not going to stand in front of their people the next morning at 6:00 and say, this is what we're going to do today. I mean, if there is one mistake in there, like, Pat, we already built that or that needs rework. Don't you notice? Back to the trust, right?
11:07 Hanne Tidnam? Yeah.
11:08 Speaker 03 Their credibility just dies. And so that's where we just couldn't keep it up, right? We were able to keep it up for two, three days and then just eventually the reality and the plan fell apart. But I'm excited to see. Is the combination, the tools that automate the planning and the sensing and the technology to abstract what really happens on the site into actionable, a managerially actionable information that can be matched against the plan. Definitely. To actually bring the daily feedback and the reality, like this work really takes this long back into the planning.
11:41 Hanne Tidnam? Oh.
11:41 Speaker 03 So that we can make better, we can increasingly make better predictions.
11:44 Hanne Tidnam? Well, there's your feedback loop, like not just on a daily level, but on a gigantic planning level.
11:48 Speaker 03 Eventually, we will have all of this data, you know, like a baseball coach has this kind of data for who to send out for pitching in various situations.
11:58 Hanne Tidnam? Yeah.
11:58 Speaker 03 But we don't have it in construction.
11:59 Hanne Tidnam? Oh, that's so funny to think about, yeah, how accessible it is in other certain enclosed finite scenarios. What are the new kinds of data that you're factoring into your decision making?
12:07 Speaker 02 We're just trying to use historical data is a big thing for us, you know. Prior it was just, here's a set of plans, go build it and on to the next one when you're done. And now it's like, how can we leverage that data to make our business better? So productivity rates and stuff like that, if we can have very accurate productivity rates, it helps us on our estimating projects for the next projects. If we're getting schedules from our subcontractors saying it's going to take two months, but in reality, we know we have this much to install with this kind of productivity rate, we can now say, well, it should actually be this.
12:31 Speaker 03 This illustrates very nicely how, in my experience, the technology actually builds trust. Everybody functions on their beliefs and the things they remember. If we have the data, we can start to gain trust in each other.
12:43 Speaker 04 One of the big challenges is just getting the right data because let's say that you know that a particular assembly takes six weeks to install, right? You have no idea at an object level, what happened in those six weeks. You've got human reports that I got 10% done, 20% done, 30%. You have no idea if that's true or not.
13:00 Hanne Tidnam? But what do you mean at an object level?
13:02 Speaker 04 Well, a construction project consists of millions and millions of objects and every one of those objects takes a different amount of time to install.
13:09 Speaker 03 Building components, parts, things like that.
13:12 Hanne Tidnam? It's reminding me of the escape room I did on Friday.
13:15 Speaker 04 Yeah.
13:15 Speaker 02 Find the key to incertitude.
13:17 Speaker 04 Exactly, exactly. Building off of that analogy, if you knew exactly how long it takes for each step in that process, and you see a new escape room, which has half of those components, but doesn't have the other half, you can now suddenly use that data because you have a history on it. On the other hand, if all you know is that, you know, escape room 11 took six weeks and escape room pirate took eight weeks, what good is that for your estimation? So a big thing in construction right now is how do we get hyper granular data on how long things are taking to get installed, how much labor it's taking to get them installed, and things like that.
13:50 Hanne Tidnam? So it's surprising that you can get so much just visually. Is it just visual, this data that Doxel is collecting, or are you also collecting? Because like, how do you know if a switch has been installed, the wiring has been connected in the wall?
14:03 Speaker 04 Yeah.
14:04 Hanne Tidnam? Like, how do you get that granular from just a vision point of view?
14:07 Speaker 04 Yeah, that's a good question. So we use vision as well as LiDAR, which tracks the 2 millimeters of accuracy. We can't yet detect things like wire pulls. Having said that, most of the issues, as well as the progress challenges, the progress management challenges in construction, take place at the MEPF stage, the Mechanical, Electrical Plumbing and Fire stage. And the reason why that happens is because multiple trades own the site and have to share it. So backstory, right? You start construction with excavation, pretty much one trade owns the entire site, right? Then you move to foundation, structural, so on and so forth. There's just one trade that owns the whole site.
14:45 Hanne Tidnam? They're running the show in that phase.
14:46 Speaker 04 They're running the show, it's just them. The moment you get to MEPF, it's this ballet, which has to work perfectly well. If I'm a day off on getting my fire sprinklers installed, I could hinder the duct guys from coming in and doing their job.
15:03 Hanne Tidnam? And is that because of efficiencies of timing largely? Or are there actual mechanical issues as well that when everything is not perfectly choreographed?
15:13 Speaker 04 Oh, both. It's like, you know, not to sound too trippy, but it's a very time and space kind of problem. If you don't install what you promised at the right time and at the right place, which could sometimes be half an inch, quarter inch accurate, you could create like a massive ripple effect. And sometimes these issues are discovered two or three weeks later. There was a project on which the Robinson deck was installed incorrectly. Someone put in the wrong deck, poured the concrete, the concrete had all set, and basically the entire floor was like wrong.
15:43 Hanne Tidnam? How often does the ripple effect get that big?
15:46 Speaker 04 So 20% of construction costs are usually rework costs.
15:49 Hanne Tidnam? Wow.
15:50 Speaker 03 It seems on most projects, there's not a single project where things in the right place land for the furniture. And then, of course, the people notice and say, well, your furniture is wrong.
15:57 Hanne Tidnam? It's so true. Even on like Minor Home, I had a friend redoing a bathroom.
16:03 Speaker 03 The alignment.
16:04 Hanne Tidnam? redoing his bathroom. Exactly. In Palm Springs, and it looked gorgeous until he sat down on the toilet and his feet couldn't touch the floor. And then he was like, okay, wait, this is weird. Something's off here. I can't imagine when you extrapolate out to projects that are hundreds of million dollars.
16:19 Speaker 03 Another key difference between manufacturing and construction. I mean, there's also many similarities. We create physical products and so on. But in manufacturing, you don't have to deal with the workspace. In construction, you have to allow for the workspace that a work crew needs, and the space is changing over time. In construction, it changes every day, the workspace you have, the workspace you create, you know, the workspace you need for many different reasons, safety, you know, just productivity and so on. So that's something that is very unique to construction. That's very challenging. And if you think back to manufacturing, what made the improvement possible there is to look at the installation of every object of every part and then to really align the design and the manufacturing methods to support each other.
17:03 Speaker 03 I just see this incredible improvement possible, not only for the construction process, but then for the synergies of design and construction that being created by having this granular data.
17:13 Speaker 02 Kind of construction is always rework. What excites me is that you can catch that a lot earlier. And so maybe if we're able to cut that 20 down into 10% because we've caught errors, you know, in a day instead of waiting two weeks to find out when the next trade is trying to install. And that can save us a lot of time, you know, cut down that rework. And so there's a lot of potential there and that's what gets us excited about that feedback loop.
17:31 Hanne Tidnam? Right. You don't necessarily even have to solve the problem perfectly. You just have to shave a little bit off that 20%.
17:38 Speaker 02 Yeah, it's getting that feedback loop to know there's an issue and then take care of it.
17:40 Hanne Tidnam? So what does this actually look like on the job site? Describe to us how this data is being collected. It's an autonomous vehicle. How does it move around the different terrain?
17:48 Speaker 04 Sure.
17:48 Hanne Tidnam? What does it need to know in this constantly shifting environment? Also, how is it making decision? Because presumably at the end of the day, you're still presented with this volume of data. I mean, how is that parseable in a moment of five to 10? I don't know how long it takes to change your plans for the next day to shave off, you know, a little bit of that percentage.
18:09 Speaker 04 So first of all, you're not presented with the volume of data. That's what the AI's job is. It's to translate the data into a digestible format. So on DPR, for example, DPR is one of the fastest moving general contractors that we've worked with. And I'm not buttering you up. That's very honest.
18:25 Hanne Tidnam? And by fastest moving, you mean to completion of a project?
18:27 Speaker 04 Schedule wise, it's just the furious pace, which we've not seen before. So usually we do like once a week reports and many contractors, as well as owners have told us, that's a little bit of overkill, right? Because we can't move the field that fast. DPR is doing two reports with us a week and they're saying, can you do it every day?
18:44 Hanne Tidnam? So what do you attribute that fast moving culture and pace to?
18:47 Speaker 02 Well, the pace is generated by our ability to plan. And so the job that Sorab is mentioning is a lab job. We modeled the entire job all the way down to, you know, three quarter inch conduits. So everything in the wall was modeled, everything above the ceiling was modeled. So that way, everything was being fabricated offsite. And we have a lot of shared rack systems that are coming with mechanical, electrical plumbing and then fire protection systems already installed on it. And so these systems come out pre manufactured. And so it's getting us more towards a manufacturing where it's more of a kit of parts that we're installing now as opposed to, you know, cutting every piece of conduit in the field and then having to measure it and install it.
19:19 Hanne Tidnam? Like Legos.
19:19 Speaker 02 Exactly. We're trying to get more towards an entire Lego set to put a building together. And so we're doing all that through the use of virtual design and construction or building information models. And that's how we're getting to that really rapid pace of getting MEP installed. And then now, once we get a feedback loop of making sure that's installed correctly, we're able to make sure we're not missing something or find those errors early in the process. So that way, if we're running fast and we have an issue on day one, you know, if we don't wait till day five to find that, it's, you know, it could have been, you know, day 15 on a normal project.
19:46 Hanne Tidnam? What are the kinds of things that you catch?
19:48 Speaker 02 It'll be things like, you know, main duct work is installed two inches off from where it should have been.
19:53 Hanne Tidnam? It is just inches a lot of the time.
19:54 Speaker 02 A lot of, especially in labs.
19:55 Speaker 03 But you only have inches.
19:56 Hanne Tidnam? Yeah.
19:57 Speaker 03 Because otherwise you have to build your building bigger. That means more structure. That means in an earthquake zone. I mean, it's just the ripple effects on on cost are just tremendous.
20:06 Speaker 02 Especially on an additive process of known as construction. You know, an inch here, an inch there adds up to two inches, which could ultimately be an issue. We have found enough that it's been valuable for our team to know exactly what's installed and when it's installed.
20:17 Speaker 04 The biggest piece, which I think seems to be grabbing the most eyeballs, is the schedule update. Construction schedules look like they're basically Gantt charts, right? They're these massive Gantt charts which are showing you dependencies, right? And what our tool does is, once these robots go in and do their thing and capture the data, the AI updates that schedule completely automatically and shows you in the same Gantt chart exactly where you stand on schedule.
20:42 Hanne Tidnam? So the general contractor isn't standing there reading a report. It's direct to schedule.
20:47 Speaker 04 Exactly. They're looking at the same schedule that they look at every week in their schedule planning meeting, except that this time it's updated with superhumanly accurate reports from the field generated by AI.
20:59 Hanne Tidnam? Is it clear why? Do you get sort of reports on like, well, this is why? Because this electrician slept late. Is it clear or is it invisible behind the scenes?
21:08 Speaker 04 So that piece really is decoded very quickly by the superintendents on the site. Our job is to tell them that there is a problem. And oftentimes, they don't even know that there's a problem. Like, there have been multiple conversations. And this is just in general, going back to the trust issue, right? Where someone says something is installed or someone thinks that something is installed and it happens to not be installed, right? Maybe there was a communication gap or whatever led to that. But it's very, very important for the top decision maker who the buck stops with on that project to know that it's not installed so that they can reorganize the field based on that. So that's basically the value of the AI, that it translates it into a readable format rather than giving you pictures or laser scans or something of that, which is just counterproductive.
21:50 Hanne Tidnam? How are you training your AI to recognize different kinds of iron beams and different kinds of flooring? Do you send it out into, is it learning as it goes? How, you know, what are the challenges behind that kind of AI?
22:04 Speaker 04 So the state of the art in computer vision prior to, say, deep learning is either one of the most popular techniques is this thing called SVM, right? Support Vector Machines. So Support Vector Machines SVM is basically this machine learning based approach, which trains a computer to recognize a certain type of object with a certain type of sensor in a certain type of environment. And it takes a smart engineer maybe two to three months to program it for something super tough. Sometimes it could take just a week, but the point is it's customized programming for every type of object. And the moment you change the sensor type or the type of environment changes, it doesn't work. So for example, if you train a system to recognize a pipe,
22:46 Speaker 04 but when you go into a different type of construction site where that pipe looks a little bit different and it's blocked and you can't see the entire pipe.
22:53 Hanne Tidnam? Right, this does not sound like it would work.
22:55 Speaker 04 Exactly. The computer would just be like, oh, this is an alien object. I've never been trained to recognize this. Deep learning was just beginning to show promise in 3D computer vision. And the value of deep learning is that you can achieve a higher layer of abstraction where it can train itself on recognizing all these different types of objects in different environments with different sensors. So you don't have to now program 20 different types of objects for every construction project.
23:23 Hanne Tidnam? Right.
23:23 Speaker 04 You've got the same neural network that's carrying over learning from construction site A to construction site B.
23:28 Hanne Tidnam? Right.
23:28 Speaker 04 And furthermore, it can do that with different sensor types. So you're not bound to the same sensor. As sensors evolve, the same neural network can adapt and can learn and can use learnings from sensor one and translate it to usage from sensor two.
23:41 Hanne Tidnam? What were some of the Chihuahua blueberry muffin challenges that you guys struggled with?
23:46 Speaker 04 I think the biggest challenge frankly was the fact that there was no publicly available work, which had shown that the same quality of recognition could be achieved with 3D data as with say 2D data. Like in 2D data, ImageNet was the state of the art.
24:02 Hanne Tidnam? So a giant reservoir of visual representation of construction sites, basically, would have been ideal.
24:09 Speaker 04 It's not just that. It's also about the core logic of how computers learn how to deal with 3D shapes, because it's completely different data. And just on the core algorithm level, there wasn't as much research out there as with say 2D computer vision. And then the second challenge was how do we get all the data? We have the highest accuracy in 3D and construction environments in any publicly available project that I know of at this point.
24:32 Hanne Tidnam? So in my unfair ratting out of the electrician sleeping late, could it feel that way on a human level? Like, do you have this, you know, sort of robot following you around, reporting on, I've done this correctly, I've done it on time?
24:45 Speaker 02 It definitely could. But if we go around just pointing fingers at people, saying you did something incorrectly, they're never going to want to work with us again. And we're going to go out of business sooner rather than later. What it is useful for is to facilitate those conversations, because if we don't catch something, then it's going to become a bigger issue down the road that could lead to, you know, in the worst case, litigation. And so if we present the data in a way that it's more collaborative of, hey, we're coming back, we found this, let's go talk about this real quick, get it fixed now. So that's not a larger issue down the road. You gain the trust of people at that point. And then also, you know, after you've created that kind of culture, people actually want to bring some of these issues to you. Or a lot of the time we talk about-
25:18 Hanne Tidnam? Oh, it's a culture shift. Yeah.
25:19 Speaker 02 Definitely. And like, we always talk about go back work in construction of, you know, the electrician installed everything but that one box because he didn't have it on site that day. And he's going to go back in three days and go install it. Usually you wouldn't catch that as something that happens, but that one box being installed could affect the HVAC guy that's installing that day. They know it's not installed, it's going to show up on our report as not installed. So instead of waiting, now they're coming to us saying, hey, I don't have this on site right now. What can we do to make that shift? And so we can have that discussion with all the other subcontractors in the room and say, hey, can the electrician get in to saw that box on this day? Is it going to cause an issue? Yes or no? And so it creates more of a trust throughout the entire job site as opposed to deteriorate as long as you use the information in the right way.
25:57 Speaker 03 It's a big benefit to find the mistakes. Absolutely. But it's actually, I think, also a benefit to document all the things that were installed correctly.
26:05 Hanne Tidnam? Ah, of course.
26:07 Speaker 03 So then you can actually get a quality score for your subcontractor. You know, we focus a lot on mistakes and we have to.
26:11 Hanne Tidnam? Yeah.
26:12 Speaker 03 But I think it's also good to know, hey, this subcontractor installs on our projects 99% of things correctly the first time. And this other one only 95% or whatever.
26:20 Hanne Tidnam? Yeah.
26:21 Speaker 03 And you can show the owner creating that trust as well, I think is important and that data. And from a business perspective, I think this technology does really two things. It improves the predictability. And the second thing is it reduces time to market. And then in the kind of technology complex projects that DPR is building, that is super important. Because if your time to market is too long, the technology will change, then you have to do rework because of that.
26:44 Hanne Tidnam? Can you give an example of what that might?
26:46 Speaker 03 So they take a hospital and they commit to a particular MRI machine. And then, you know, it takes a long time to build the down thing. And by the time, you know, six months before the hospital turned over, a new one comes on the market. And like, they don't want to open the hospital with an MRI machine from three years ago.
27:01 Hanne Tidnam? Right.
27:01 Speaker 03 But then that one will have different foundational requirements, different electrical connections, et cetera. And you are having a lot of fun with.
27:10 Hanne Tidnam? Trying to manage that.
27:11 Speaker 03 And so if you can shrink that time to market, then the client still gets what they ordered in a timeframe that's useful to them.
27:17 Speaker 04 If you look at anything in the world, trust issues exist because two people have different versions of the same facts. Right? If both parties trust a third party to come up with that version of facts, which they believe is more objective, it inherently increases trust if both parties begin trusting that. Not just for general contractors and owners, but also for subcontractors. Because if you actually look at how subcontractors get affected by this lack of trust, they often get affected in the form of cash flow. A subcontractor says, hey, Hannah, I'm 25% done. Could you pay me 25% of my budget of what you owe me? And Hannah is thinking, damn, if this guy is 24.9% done and there's like one pipe installed in the wrong place,
28:00 Speaker 04 I'm going to be liable for a three week schedule delay when I discover it four weeks later.
28:03 Hanne Tidnam? Oh.
28:04 Speaker 04 So Hannah doesn't pay Mr. Subcontractor.
28:06 Hanne Tidnam? Despite the fact that it's.
28:07 Speaker 04 Despite the fact that the sub is claiming.
28:09 Hanne Tidnam? And then trust is eroded again.
28:11 Speaker 04 Exactly. Now, instead, if Mr. Subcontractor comes up and says, I have this two millimeter accurate laser scan and the world's most sophisticated artificial intelligence. Shameless plug, saying that I'm 24.86% complete, could you pay me? It gets a lot more objective. So it goes far beyond just the job itself. It starts affecting project financiers. It starts affecting project insurers who get more transparency on the job.
28:38 Speaker 03 We see just all of these wasted efforts, frankly, from many, many players, from subs to GCs to owners, designers. We see all these wasted efforts that will go away. With AI, I can do more things more consistently than I'm already doing. So that is, let's say, in this case, apply a particular inspection routine more consistently. Then I can do it better because I can, over time, realize, okay, this works 99.8% of the time.
29:13 Hanne Tidnam? The feedback loop.
29:14 Speaker 03 And then I can do new things that I wasn't able to do otherwise because I can learn. I can, you know, we cannot look across data from 50 projects and make sense of it. That's super cool.
29:23 Speaker 02 From the construction side, you know, one of my goals is to make our teams more efficient and to get our teams and our people solving problems instead of doing non-value-add tasks. And going around and looking at a model to compare it against what was installed isn't necessarily a value-add task. If we can automate that, we can get the people that would be doing that job out in the field, solving more problems, working through more changes, making sure the owner's getting what they want, you know, all those kinds of things that would trickle down from there. And so that, you know, for me, that's a huge benefit.
29:48 Hanne Tidnam? Maybe we can quickly touch on the regulation. Is there, you know, regulation is notoriously difficult around construction sites and introducing new technologies, even down to, we talked in the other podcast about like even when you wear your glasses and what it's like to bring an iPad into the field. What's it like to bring a autonomous robot into the field? How does the regulation work around that?
30:09 Speaker 04 As far as the actual robotic equipment that we're bringing in, the hardware is nothing new. We're using sensors that have been used on construction sites in the past. We're using things like cameras. So it's not rocket science. The hardware isn't particularly novel. It's pretty much off the shelf. So from that perspective, you know, from a regulatory standpoint, we haven't had as much of an issue. Of course, we also capture data with drones. So there, of course, you have a standard process, which is, you know, FAA Part 107, and you have to have a certified pilot and things like that. But other than that, you know, we've not faced as much of any challenge.
30:45 Hanne Tidnam? When you use the drones, is using the airspace of a construction site, isn't that a challenging thing in terms of also coordination and timing?
30:53 Speaker 04 Yeah, it is actually. FAA guidelines don't allow you to fly over anyone who's not in direct operation of the drone. So we only do that after hours.
31:01 Hanne Tidnam? Oh, okay.
31:02 Speaker 04 Now, having said that, it's a very little known fact, but 80% of construction money is spent indoors anyway. So the 20% that's captured with drones just goes by super fast.
31:12 Hanne Tidnam? Well, that makes sense, actually, when you think about what a building is like.
31:15 Speaker 04 Yeah, exactly. Now, to touch upon the point you made about regulatory authorities in construction, it's actually a huge deal. And, you know, we do a lot of work with healthcare, which is one of the most regulated construction spaces around. And a lot of the events in the construction site are actually planned around when the inspector will show up. My kind of dream, be able to have as much trust in our technology as general contractors and owners do, to have that same level of trust, you know, exist with the inspectors. Because now, all of a sudden, they can remotely take a look at what's getting installed, what's not getting installed, trust the quality that things are getting installed by, ensure that it is indeed the plan that they had stamped off on.
31:54 Hanne Tidnam? So you don't feel like you're showing up and you're going to trial one day? It's just there's a constant flow of information and you've been doing things right all along?
32:00 Speaker 04 It's just about the logistics of managing an inspector showing up on site. I mean, every site that I go to, there's a board that says, "Reserve for Inspector." Like, that's a reserved parking spot for the inspector, right? Because they show up on site and then sometimes you'd have to have some parts of your wall exposed.
32:18 Speaker 02 It's usually right next to the red carpet.
32:20 Speaker 04 I'm sure sometimes you guys don't close up walls and stuff so that inspectors can come in and take a look at it.
32:25 Speaker 02 Especially the Office of Statewide Health Planning and Development work. They're the most stringent regulatory agency, I think, in the world, based on the seismic requirements for the state of California. That's why building hospitals has been so expensive in California for so long.
32:38 Speaker 03 Because we actually want to use them after an earthquake.
32:40 Speaker 04 Now, imagine if, you know, OSHPORT trusted tools like this and just said, "Okay, yeah, cool, you can cover it up, right? Just keep doing your work. We have a system of record which is showing us this stuff got done correctly."
32:50 Hanne Tidnam? Gathering this kind of information must feel like a goldmine is opening up. What are the kinds of questions that this kind of new data makes you want to ask? If we go way past, like, how efficient is this job site, and we turn all these job sites not into little labs by day, but like, multiplied by tens or by hundreds, what are the kinds of questions, the big questions, that we might start to get answered?
33:11 Speaker 03 I think from a societal perspective... If you can cut 20% of the rework and then you could make construction 20% more efficient in other ways, which seems entirely within the realm of feasibility, you're talking improving construction by 40%. In most economies, let's say a rough number is construction represents 10% of GDP. So imagine freeing up 4% of GDP to do something else.
33:37 Hanne Tidnam? Yeah, that's a powerful new resource.
33:39 Speaker 03 It's crazy what you could do with that amount of money from there. Now, from a project and company perspective, I think being able to connect the business purpose of hospital, fabrication plant, whatever is being built, and translating that into actions that take place, strategic actions that you have to do every day on the project to meet that business purpose. I mean, that's a connection we don't have today. And then those are, I think, the kinds of questions we can ask. What kinds of designs will enable this kind of production or what kind of flexibility you seek maybe? So connecting really the client purpose with design and construction and aligning those, that's where I see the foundation being created.
34:20 Hanne Tidnam? It actually, it's interesting that it sounds like an opportunity to rethink a lot of the foundational assumptions we've been making for a long time.
34:27 Speaker 03 I think once we have these feedback loops, we will have a construction industry that we haven't known so far.
34:31 Speaker 02 The one thing I would add is be able to leverage this data from a construction standpoint and applying that data that we find, such as productivity rates or how long, you know, procurement lead times, et cetera, into objects that we then use for design purposes. And that creates an even bigger design loop of now if we provide construction objects to a design firm to use in their design, now we know what the cost is, what the schedule is associated, or productivity rates, whatever it's going to be. And that way, we can almost start utilizing that information right off the bat to start building estimates and schedules with the design information. So like when an architect or an MEP engineer is going through to create their design, and whether it's, you know, a linear photoduct work or if it's a door, if they can populate a door that knows, you know, the lead time on this door is six weeks, you know, the frame is seven weeks, and if we need to have that door there in two weeks, we should probably not be using that object.
35:16 Speaker 02 And so if we can provide that kind of information upfront in the design, there could be a lot of benefits there. Or if an architect is starting to create their design, they drag a wall and see that that wall is, you know, $100, $200, $300, $400.
35:27 Hanne Tidnam? All that context is right there.
35:29 Speaker 02 And now they can start using cost information to help create their design as opposed to create a design, wait for an estimate, go back and revise the design to try and meet an estimate, you know. There's a lot of that back and forth that could be eliminated if we start providing that real time feedback with this giant feedback loop of
35:44 Hanne Tidnam? Changes decision making from day one.
35:46 Speaker 03 We did actually a test of that for steel structure. We were able to get the data pulled together and we were able to cut the cycle for feedback from eight weeks that you have today in terms of creating a design, you know, getting the estimates for fabrication, et cetera, and getting it back to a few hours. I think today it could be minutes and reduce the cost of the structural by 13% and make it 20% faster to build.
36:11 Speaker 04 I think what both of these guys are talking about is fundamentally how data captured in the field is going to make planning even better. Every time a construction design is built, the estimate of how much something costs and how much each object costs is completely fresh. Ultimately, any construction project is a trade off between budget, schedule and quality. So as you start getting these insights from the field, it just becomes easier and easier where you can optimize better for your owner. Like, maybe it's an owner that doesn't care as much about budget, but cares a lot about schedule and quality.
36:42 Hanne Tidnam? So to go from design to reality to how you measure reality back to design again.
36:47 Speaker 04 Exactly.
36:48 Hanne Tidnam? Yeah.
36:49 Speaker 04 Yeah.
36:49 Hanne Tidnam? That's a big impact. Well, thank you guys so much for joining us on the A16Z podcast.
36:53 Speaker 03 Thank you.
36:54 Speaker 04 Thank you.
Previous Company Building in Crypto: A Conversation between Brian Armstrong and Chris Dixon Next What's Next for Education Startups Page updated October 9, 2026 · transcript and summary are machine-generated