Bonus Episode: Trunk Tools CEO Sarah Buchner is Building Trunk Tools to Simplify the $13T Construction Industry

Sarah Buchner is the Founder and CEO of Trunk Tools (https://trunktools.com/). She is a one-of-a-kind founder, having spent her young life as a carpenter in Austria and then working her way up the ranks of the construction industry. She’s also earned several graduate degrees including an MS in Civil Eng, a PHD in Data Science and an MBA from Stanford. Trunk Tools is an AI Tool for the $13T construction industry aimed at enhancing project management and addressing the skilled Labor shortage. We wanted to have Sarah on to share her experience in building a vertical AI tool - specifically: where AI will have the biggest impact on construction and behind the scenes of how they’ve built tools like TrunkText and TrunkScheduler to massively reduce the amount of rework in construction.

• 0:00Speaker 00

  I'm Mira Clark, an early-stage investor focused on vertical SaaS, consumer and prosumer AI tools here at Redpoint. Today, we have a special episode of Unsupervised Learning with Dr. Sarah Buchner, the founder and CEO of Trunk Tools. Trunk Tools is one of Redpoint's latest early-stage AI investments, and we recently announced their $20 million Series A, which we had the privilege of leading. Trunk Tools is reshaping the world of construction, a $10 trillion industry that is defined by masses of unstructured data. At Redpoint, we see this overwhelming amount of data as a prime opportunity to leverage generative AI to drive greater efficiency moving forward. Since launching their AI product in the past 12 months, Trunk Tools has onboarded many of the largest construction firms in the country as customers and is already delivering seven figures in cost savings to these customers a year.

• 0:45Speaker 00

  Sarah, as you will pick up on from our conversation today, is a true outlier of a vertical SaaS founder, having entered the field of construction at age 12 as a carpenter before later going on to earn both an MBA from Stanford and a PhD in civil systems engineering and data science. She's an absolute force and has certainly recruited a team of similar caliber around her. In keeping with the theme of unsupervised learning, we asked Sarah about her biggest insights in building in vertical AI today, as well as her hottest takes about the future of construction. You'll hear that conversation with Sarah now. Sarah, thanks so much for being here and so happy we were able to catch you while you're on the West Coast.

• 1:20Speaker 01

  Thanks for having me.

• 1:21Speaker 00

  So, to dive right in, clearly a lot of opportunity across the board, but in the next decade, where do you think AI will have the greatest impact on construction?

• 1:30Speaker 01

  Mostly when it comes to data, right? And I think it's so interesting when I did my PhD about 10 years ago, and I did it with data science, AI and civil engineering. And it was just like data science in construction was not a thing at all because data was there, but data was terrible. So, many years ago, for me to even do research, I had to spend tons of money to even structuring data so we could do anything with it. And there has always been a problem. Over the last, I would say, like five to 10 years, large general contractors or large companies realized that and they started structuring the data and data warehouses and data lakes. And, you know, the McKinsey's of the world came in and... For the big companies, that was something they could afford and could do. The mid-market and the SMBs were just lost, right?

• 2:11Speaker 01

  And what's interesting now is that generative AI is so good at structuring, not so good, but it's decently good at structuring unstructured data. And then actually let people give them insight into what's going on with the data, which is now accessible to the whole market and not just to like large corporations anymore. They're throwing tons of money at it. So I think that's one thing. And then the second thing is just, um, construction is incredibly bureaucratic and it became more and more bureaucratic over the last, whatever, three, four decades, because if it's a very litigious industry. And so if it's litigious, people start, you know, documenting everything. And if you document everything, you have a ton of unstructured data, which again adds to the first problem.

• 2:51Speaker 01

  So I really think generative AI is touching both of these foundational problems in construction. And that's why I think it's taking off so much.

• 2:59Speaker 00

  A hundred percent. I guess as you think about both the ability to better leverage that unstructured data, as well as this whole empowering workers to do the work they want to do. Can you set the stage for our listener? Like, what is Trunk Tools? How does it solve that?

• 3:13Speaker 01

  So we are a platform that builds the brain behind construction. That's how we try to explain it in a very easy way. What it is for the more, you know, AI nerds out there. It's a system that goes in and takes all of the unstructured data that is existing in construction and structures it in different ways. And then on top of it, we build an army of AI-based agents. These agents, it's like somewhere between like rational and intelligent agent, depending on what the actual task at hand is. But they are fueled by this brain that has structured information in it. And then are acting on behalf of the user and actually take over full workflows. Why this is particularly interesting in construction is because for an average high-rise project,

• 3:57Speaker 01

  let's say like a $500 million high-rise, which in construction that's average. For an average high-rise project, you have about four million pages of documentation. So the contract is four million pages long. It's the longest contract in the world. And it's, you know, it's a multimodal contract in a way. It has pictures, it has blueprints, it has schedules, it has tables. And so it's a complete mess of unstructured data. And the way we actually attract talent is we tell people like, do you want to play with the messiest dataset of your life? And that's why our AI engineers are so happy, because they get to play with like real problems in the real world.

• 4:31Speaker 00

  You're building an app. It's not a model. And so as you sort of think about underlying model selection then, like what does that evaluation process look like for your team?

• 4:41Speaker 01

  If there's a new model coming out, we test it specifically on each of one of our problem sets and use cases. And if we like it and we can, then we usually negotiate, we get an enterprise agreement in place. Or, you know, we try to make it safe, not try to, we're making it safe so that actually our customer data is safe again. And once it passed the testing and the data and security side of things, and we actually think it's good enough, then we swap out whatever we had there before. I think it's a big advantage towards incumbents who have just decided early on to partner with one and only one model provider, or one LLM, or one visual model, or whatever it is. Because we actually benefit from the fast speed of development in the whole world.

• 5:25Speaker 01

  And we're not just like stuck with one partner. And so like our performance of our overall products will always be the best in the world because we just plug in whatever best technology we find on top of everything that we have already built.

• 5:38Speaker 00

  Yeah, I guess as you sort of think about solving those problems then, for the customer, are there any dynamics or capabilities that you're really waiting for with these underlying models to better serve the customer?

• 5:50Speaker 01

  We are waiting for a little bit for like speed of response to increase. The faster we can get information back, the more complex work we can do to build on top of what the underlying models and what we have right now. It's very interesting because construction is very complex, right? I said this before with like the four million pages, but that's not the only thing. Actually, construction is a huge knowledge graph when you think about it. And you have, the real problems in construction come from the third layer effects of a change, right? Like you intellectually, as a human, understand if a change A, B, C, D is happening. But then being able to actually understand the next layer and the next layer that has impact of millions of dollars and, you know, like weeks and weeks and months and months of delay,

• 6:35Speaker 01

  that is really complex. And we Cannot fully crack the layer yet because again like the the the speed of response is not there. Um, and and just some of the yeah, some of it's also a little bit cost prohibitive in some ways around vision models. But that's all stuff that we think it will go down regardless.

• 6:52Speaker 00

  On that cost component, where has Empress Cost played the greatest impact on how you've approached building the platform?

• 7:00Speaker 01

  We do not really care that much right now because the problems we are solving are so big and so impactful that we don't have the problem consumer products do, right? Like, it's like, oh, my God, can I charge $20 a month? Like, we are saving hundreds of thousands of dollars per construction site.

• 7:16Speaker 00

  Well, we've heard you're saving millions, but let's call it hundreds of thousands.

• 7:19Speaker 01

  Right, like, so like, again, yeah, like, we're saving like hundreds of thousands per per job site. It's like, yes, in the overall economy, we're saving like so much money that it it really doesn't, it doesn't matter too much. Um, I think where we haven't, where we are not fully processing everything is around photos right now on construction sites. Like, we've run a lot of tests, it's really exciting, but it's the D R O I purely on on images is is not as as high as as we have in other other parts of our business.

• 7:48Speaker 00

  I guess as you sort of think about how you all have approached fine tuning the model specifically for construction jobs, like, are there insights or learnings that can be abstracted away for other verticals?

• 8:00Speaker 01

  I personally have no idea about any other industry. Like, all I've spent my life on is like construction and tech, so that's all I know. But no jokes aside, I I think there's lots of learnings. We definitely have a very, we have a bunch of industry specific data sets and file types, which I'm sure other industries do do as well. And the reason why our customers are saying that we're so much more accurate is really because we have written custom parses for our industry. You know, we've fine tuned these models and we've done a lot to really bring the accuracy up for the specific industry. Even just like. Translating approvations and words and like lingo, it's very basic stuff, actually. I'm sure it's the same in every other industry as well.

• 8:43Speaker 01

  And then eventually you build a data set that is, you know, very, very interesting for, again, in our case, ten percent of GDP is driven by the data set that we are creating or have created. That becomes really powerful. So I'm sure you can do it in other industries. I think to my knowledge, there's only healthcare, which is a bigger industry than construction, and a lot of people are doing this, but I think find your little niche. And even if it's just like one percent of GDP, like you have a huge market.

• 9:09Speaker 00

  I mean, the data set dynamic is super interesting, right? Because it only compounds in terms of benefits over time. Today, though, you sort of sit on top of a lot of these existing systems. And so it's a stack which you're part of, but you're not necessarily like their day one foundation. Over time, as your feature set continues to expand, like, how do you think about a desire to become the primary provider in the stack over time versus kind of staying in X, Y, Z lane and continuing to run full speed ahead there?

• 9:40Speaker 01

  I think we are on this interesting cusp of of change in technology in general, right? I think, and that's another polarizing view, but like if I look back at the last ten, twenty years of software development or like SaaS, it was always like cloud storage with some predefined workflows on top. That's what most, at least in construction, it's most of these products are. And, you know, that was the right move. I think where this can go right now is like you have a cloud storage where you dump your unstructured data. You have AI structuring as as much as possible, which is what we call our brain and the brain behind construction. But then on top of it, you can actually have an agentic architecture that just goes in and creates these workflows on the fly. It's a very different dynamic. That's what we have right now.

• 10:22Speaker 01

  What we are not, what we are not right now is the original source of truth from a data perspective. But that's the easy part about it. So what we are focusing right now is really gaining like traction and market shares purely and getting people used to letting AI agents work with them or for them or whatever you want to call it. And we did a big analysis and showed that about seventy percent of the Of the bureaucratic workflows in constructions, or not the person hammering, but like everything else that's happening in our industry. Seventy percent of those workflows can be automated with the tech that we have today.

• 11:00Speaker 00

  Which is insane.

• 11:01Speaker 01

  That doesn't mean it's productized, right? Like, just to make sure, but the technology is there. So all it takes now is companies like us going in and actually productizing this so it's useful. And obviously with the quick, you know, the speed that you can deploy these AI agents and develop them, that will happen really, really fast. So our goal for for the industry is more to become like the best friend, the best buddy, the the the buddy that is out there in the field, taking off all the work that you don't want to do regardless. And once you have that, changing the rest of the story is is the easy way. So that's how how we think about it. But right now, I have no interest in becoming like the project management system.

• 11:39Speaker 01

  I want to just become the best buddy in the field, the number one interface, a semantic way of helping somebody help you do your work.

• 11:47Speaker 00

  Yeah, yeah, trunk tools, best buddy, I feel like you like alliteration. I guess maybe double clicking on that best buddy point. A lot of it, as you said before, kind of comes from the fact that a lot of your team comes from industry. And I think with a lot of these vertical AI applications, we've seen teams either heavily coming from like a research oriented background or more of that industry specific domain. Kind of falling in the ladder camp, you obviously have been able to build out a pretty robust kind of tech team. That said, what have been kind of like the biggest challenges building an AI native product as a team that is not a bunch of PhDs from DeepMind?

• 12:23Speaker 01

  Yeah, we have a bunch of PhDs now.

• 12:26Speaker 00

  Yourself included?

• 12:27Speaker 01

  Yeah, exactly.

• 12:29Speaker 01

  It's interesting, I I feel like we had real troubles initially just hiring a bunch of PhDs from DeepMind, right? Because we were considered a thin chat GPT wrapper, or we were considered application AI, and nobody cared about it. And so I think it was very interesting to see how it has changed, how perception has changed and how the market has changed. We were now able to hire insanely top talent, right? And so like, and that is, I think, A, because we have a lot of traction. But I think it's B, because people understood. This recent hype of AI has three parts, right? There's like the foundational layer where like, you know, you can go to OpenAI or Anthropic or whatever.

• 13:13Speaker 01

  Then there is like the infrastructure layer and then there's the application layer. If I am an engineer, I probably don't want to be in the foundational layer because my equity is not gonna like, you know, a hundred X anymore. Like, that's kind of done. And also like, I'm gonna be such a small piece in a very like big pond. In the infrastructure layer, lots of competition, very unclear. I mean, you you'd invest in a room, right? Like, it's it's gonna be interesting to see what's gonna happen. Um, but again, it's all horizontal play. And then all of a sudden, like, the vertical application AI that we looked down upon so much in twenty twenty three is now like the shiny star because people understand.

• 13:50Speaker 00

  Isn't it so funny how that plays out?

• 13:51Speaker 01

  It's so funny, yeah.

• 13:52Speaker 00

  How do you think about resourcing or product prioritization? Or like, said another way, what are you investing in now versus later?

• 14:00Speaker 01

  We have changed, right? I mean, I wouldn't say like we've completely pivoted a million times, but you have to just stay agile in an environment where the underlying technology changes so fast, right? And where, again, we we are by far not a chat GPT wrapper, but if we were like, I would be scared shitless because then you like actually are dying. And so I think ruthless prioritization is a really important part of our company. And the way we do this is about fifty percent of our focus is on the low hanging fruits to gain traction, to gain market shares, to like gain brand recognition, to, you know, follow kind of like the Procore and PlanGrid example of the world in construction.

• 14:38Speaker 00

  And what would an example of that be?

• 14:40Speaker 01

  It's like some of the agents that we're deploying are like just really, really simple, right? Like, it's we have a meeting prep agent that goes in and prepares you from a data perspective for the next meeting that you have. Like, that is so simple.

• 14:53Speaker 00

  I feel like I need that. We generally like to wrap things up with a fun little round of rapid fire, if that works for you.

• 15:00Speaker 01

  Mm hmm.

• 15:00Speaker 00

  Okay, to kick things off, what is overhyped and underhyped in vertical-specific AI apps today?

• 15:06Speaker 01

  I love this. I think legal docs are very overhyped. Anything in the legal space, sorry, everybody out there, I had to ask me this question. I'm not an expert, but I feel LLMs are really good in like doing stuff with text-based stuff. And like, a text-based contract is just text, right? And so I feel that's very, it's very overhyped and it's also very small market. And I think like underhyped is actual workflow automation. It's really like rethinking from A to Z of like, how can you actively take away workload of humans? And how can you actively automate knowledge work?

• 15:46Speaker 00

  Which AI startup outside of your space are you most excited about right now?

• 15:50Speaker 01

  Langfuse.

• 15:52Speaker 00

  What does Langfuse do for our listeners?

• 15:54Speaker 01

  They're a great system for monitoring and analyzing language model use. And so we are self-hosting right now their open-source system, and it's been incredibly valuable. They also provide hosted options for those interested, so it's a really great tech.

• 16:07Speaker 00

  We love it. What's been the biggest surprise in building Trunk Tools?

• 16:12Speaker 01

  I think the biggest surprise was that me and also on board, actually, they were very worried about accuracy and about us having the same issues as self-driving cars, where it's like, they're stated, it's self-driving cars are supposedly better than human drivers, but we're still like too scared to actually let them go wild on the streets, because they would make one mistake and they're out, right? And so we were worried that it would happen to us, that we would lose trust by like, you know, not giving a complete answer or something like that. And it was very surprising that

• 16:46Speaker 01

  People, including the users themselves, assume that they are way worse than the AI is. And so we did some questionnaires and we're like, how often do you think you are right in answering these questions? So you would find the right information? And they gave us answers between sixty and seventy percent. And so it was like, okay, our level of accuracy right now is over ninety percent, like, great, like, we got this. And so I think that was a big surprise of everybody. Everybody was pushing us originally towards like a ninety-five, ninety-eight percent accuracy. And we, we actually, I stopped all further development at about ninety percent because it was like, we're good enough. And we're not hallucinating, right? We almost never, but we tell the users like, hey, we don't know.

• 17:29Speaker 01

  And so we have a rack product that either tells them a right answer or tells them, I don't know. And people are totally okay with the, I don't know.

• 17:35Speaker 00

  Hard question, because I know you think construction and tech and only construction and tech all day, every day. But if you could build a vertical-specific software app in any sector except for construction, what would it be?

• 17:48Speaker 01

  I, I, again, I don't understand any other market, but I, I spent some time at one point in my life writing a dating book. And so I would probably write an AI-specific dating app. Or, I'm, that might exist, by the way, I've, I've been out of the dating game for a while, but I would probably do like, train an AI agent based on who I am as a person and have my AI agent go out and meet other people. And then if these agents decide that we're a good match, actually, like, you know, have us go on a date. That's, that's stuff I would do, probably, like, play around with that. What I hate about dating apps is like, to make revenue, you actually don't want people to fall in love. And that's, wouldn't work. And that's why I'm not building dating apps.

• 18:28Speaker 00

  Yeah, I will say, this is so not rapid fire, but there's a lot happening in AI dating at the moment. And the bounties or like, you know, success prizes that people are willing to pay for a successful map, maps rather, are unlike anything you've ever seen. They're like the size of your enterprise sales contracts. So, interesting market, I agree. So, you personally started your career as a tradesman at a very young age. Yet, we keep hearing about labor shortages in the field, particularly when it comes to younger workers. So I'm curious. Does AI make the industry more appealing or less appealing to the next generation of field workers?

• 19:06Speaker 01

  I think, so here's a half serious answer. I think it's going to make it more appealing because I feel a lot of the knowledge work and the knowledge jobs are going to disappear. And so that means there will probably be more people available in the job market. Right now, the biggest problem in construction is still the labor shortage. We just, for every one person coming into the trades, five retire. It's just getting worse and worse. And we have hundreds of thousands of open jobs in construction in the US alone. And so I think the increase in labor shortage means an increase in salaries for tradespeople. And so I think that, and a decrease in salaries for knowledge workers, because AI can do a lot of that stuff. So I feel that's why it becomes more attractive over time.

• 19:51Speaker 01

  But also, AI is taking away some of the annoying parts of the job, some of the bureaucratic parts of the job. People who go into the trades want to work with the hands. And so let them go back to work with the hands. Let AI do the rest. And so I think that, combined with robotics and exoskeletons and everything else that's happening, I think makes it ultimately more attractive. And lots of jobs will change, but the trades jobs will probably almost not change. They will still be around. AI cannot actually, you know, pour concrete.

• 20:22Speaker 00

  Yeah. Finally, where can folks go to learn more about you and the incredible work you're doing at Trunk Tools?

• 20:28Speaker 01

  If you want to directly contact me, LinkedIn is probably the easiest. And otherwise, go to trunktools.com. Go to our career page. We are hiring all the time. And we need talent who is interested enough in listening to a podcast like this. Yeah, just reach out to me or reach out on our website.

• 20:45Speaker 00

  AI talent, trunktools.com. Sarah, thanks so much for joining us.

• 20:48Speaker 01

  Thank you.