In episode 4 of Hidden Layers, Ron Green, sits down with Jean Pierre Trou, CEO and Co-founder of mbue, a company building artificial intelligence systems to instantly review architectural drawings. Jean Pierre is an Architect by training and talks about pain points for all architects, the future of AI in architecture, his vision for mbue, and much more.
Ron Green: Welcome to Hidden Layers, where we explore the people and the tech behind artificial intelligence. I'm your host, Ron Green, and I'm excited to have Jean Pierre Trou join me today to talk about how architecture is being changed by artificial intelligence. Jean Pierre is a Peruvian -American tech entrepreneur and award -winning architect. He's also one of my co -founders and the CEO of mbue, a company building artificial intelligence systems to instantly review architectural drawings. Jean Pierre is the founding principal of Runa Workshop and an award -winning architecture and design field. Welcome, Jean Pierre. Happy to be here, man. Well, let's kick things off with the question about, how did this get started?
Ron Green: So Jean Pierre, can you maybe talk to us a little bit about what inspired you to want to find a way to leverage AI in the architectural process?
Jean Pierre Trou: Yeah, look, this is very personal to me. As a founding principal of an architecture firm in Austin, I found myself spending countless late nights reviewing thousands of drawings. This is me on the dining room table, redlining drawings, which is predominantly our quality control process, as is it today. And this is not a problem that is mine, per se, it's the whole industry.
Ron Green: So you see you have these large documents, just setting on your table. And I'm literally redlining. And what are you looking for in this process? A lot of things.
Jean Pierre Trou: So this is the thing, let's step back. What are these drawings, right? An architect, people think that the product of the architect is those beautiful designs that you see. The matter of fact is, we as architect engineers, we provide what it's called, set off construction documents, for you can accurately price, permit, and build a project from. So that's what our clients are paying us to procure those documents. Those documents are our product and we have to review them for co -compliance, making sure that all the different disciplines and approaches you can have the architect, the structural engineer, the mechanical, electrical, plumbing, civil, you name them. All of those disciplines need to be coordinated to ensure that actually the drawing itself is drawn correctly, you know, that the detail is correct, correctly referenced, or the information on the drawings are accurately depicted. And that process still today, unfortunately, is...
Ron Green: manual. And why isn't it a part of the process when you're laying out the design on the computer? Like why does it have to wait until a later stage?
Jean Pierre Trou: fantastic question. So we do a lot of our drawings on a 3D model called BIM, Building Information Modeling. And you know the whole process, like we can spend a whole hour just talking about BIM, but the reality is that although it's such an amazing idea of having all this information three -dimensionally, the reality is that those documents are being used to represent the end result which is the drawing, which is the you know the sheet where we built from. As of today we don't build from 3D models, we build from the contract documents which legally binding, right? And those have to go through a very strict review process to ensure that everything is coordinated and everything is per code and everything is according to the set of programmatic specifications of the project itself.
Ron Green: Okay, okay. And so what part about that process is probably like most difficult and how are you looking to use AI to streamline it or automate it? Yeah.
Jean Pierre Trou: So the problem is extremely complex because it's predominantly perceptual, it's vision, right? So you can have a symbol in your drawing that depicts a table and it's referenced as a rectangle on your drawing. So if you've never studied architecture, you looked at it, the funny thing is you probably picked that it's a table. Why? Because you see a lot of things around it like chairs and the label is a kitchen and you start saying, oh, that is the kitchen island or if it's around a room that is called dining room, oh, that square is the dining room table. So there's that association that is those relationships that are easily for humans to solve. It's very complex for a computer to do it, right? But the technology nowadays is there and that's what really led me to build an amazing team and be able to solve this for not only me but for the whole industry, you know, to generate flawless architecture.
Ron Green: Okay, so that's great. That's a great setup. So let's talk specifically about the AI a little bit more so people can get a sense of How it's going about the process so it sounds like it's almost entirely a visual process You're looking humans are looking at documents. They're looking for Anomalies like they're looking for things that are missing or things that are wrong.
Jean Pierre Trou: This is a fantastic question. The way I describe it is what I call the vertical solid quality control. So you have the graphics and annotation aspect of it that you can check to ensure that all the drawings are drawn correctly and all the information on the drawings are correctly depicted. You have the change control. When you make changes, to make sure that those changes are being picked, not only by other disciplines, but within your floor plan, you make a change on the floor plan. Let's make sure that this section, applicable to that location, has been updated, or the elevation, or anything that has been impacted by that change. You have cross -coordination. I mentioned all the disciplines that we have on the project. All of those need to be coordinated to ensure that what you're building actually doesn't have conflicts, which is the number one issue on the job side. It's the lack of coordination between disciplines. To code compliance, ensure that what you've drawn actually meets governing codes. So it's very, very complex.
Ron Green: How big are the documents? Like how many pages are we talking about on average?
Jean Pierre Trou: It depends. A house could be 10 to 30 pages, a small house to an office building could be hundreds, a couple hundred pages, or a hospital or a large development could be hundreds of pages.
Ron Green: And if it is hundreds of pages, how long does that review process typically take?
Jean Pierre Trou: I did my last review at Runa, it was a office building, 90 ,000 square foot office building, three stories. That took me a whole week to review, a whole week. Oh, wow. Yes. Okay, that's a lot longer than that. It's a lot of pages.
Ron Green: Wow, okay. So when these AI systems are helping with the review process, what are they actually doing on these documents? Well, different things.
Jean Pierre Trou: So when we track changes, what imbue does today is it can tell you with high precision what has been added, deleted, or modified on a drawing. This can be an object on the drawing that has shifted, and you have to determine if it actually has changed, like it could be a notation, it could be a dimension, or it could be a graphical change. But a note or a dimension can move in the space but not change, and those complexities are easy for us as humans to look and recognize, but for a computer it's not. So imbue, what it does is it understands when an object has moved but not changed, right? And then we can also segment the segmentation models that can basically cut, you know, in lame terms, a portion of the drawing, find its previous version, wherever it is, on the sheet, and gives you a precise description. But not only that, we're also a lot of changes happening in text, and this is critical because one character could have huge implications, like a change of a slope could have implications on water lines or gravity lines not connecting, you know, or now the elevations on this lab might differ from architectural or to civil, or create a lot of havoc. But that precision we do with a natural language processing model, where we can parse all the text from the drawings and tell you to that character level what has changed. So it changed from this to that. There's no tool out there right now that can do that.
Ron Green: And before humans were having to read the text and look for differences.
Jean Pierre Trou: Imagine you ask me how many pages, you know, a big office, 200 pages, let's say that you have to flip and read every single note because some notes may refer to, refer to a structural that means that there's additional information on the structural drawing to be found. And then you have to check if it is there and if it is not, you know, or a detail, you're drawing the same detail. This is so weird, right? So the architect draws a detail and that detail has structural components. Then additional information needs to be found in the structural. So if the structural drawing doesn't match perfectly the architecture, then there's deviations. And that's where mistakes happen. And this happens throughout the whole drawing set. So, and you catch us those, catches us on paper before it becomes a bigger issue on the job site.
Ron Green: Your company is looking at automating this sort of change management process. Are there other aspects to architecture that you're thinking about AI affecting, or do you think other companies might be looking at leveraging AI?
Jean Pierre Trou: Oh, there's a lot of workflows that can be optimized by AI. I mean, just thinking of consTrueection. The construction side has seen a lot of startups using computer vision to track shipments and material coming on site to identify what has been delivered to all the way to safety on the job side. We ensure that everyone is wearing hard hats or away from areas of high risk. So we can imagine a job site can be multiple people. Everyone is doing a specific job. There's noise and stuff. So there's really no way for you to make sure that everyone is being safe. So leveraging computer vision and that and giving notification to the foreman, hey, somebody is on this area. It could be a life saving right there. To on the design process, generative AI. That's where my mind was going. Which is, I think, everyone's mind. Generative AI. It's pretty hard right now and I've seen amazing things. I think what everyone's excited about sometime soon, prompt to model generator. Right now we've seen a lot of prompt to image generators to create design concepts. Which, I mean, phenomenal tools out there. And I feel like there is a gap between, okay, I have that concept now. I still need to put it on the model, model it, and create drawings. So I think the next step will be prompt to model.
Ron Green: Is that something you're thinking about?
Jean Pierre Trou: Oh, I love that and it's basically on our pipeline too. What I'm envisioning is throughout the whole process, optimize those areas that are the most tedious and really the ones that we are not good at. And actually empower architects and builders to focus on their most impactful work. Which for me, when I started was design. Be able to provide better design for my clients and provide value where other people can't. So that's a big differentiator.
Ron Green: So what is the long -term vision for Mbue? Like what do you see these AI systems being capable of long term?
Jean Pierre Trou: We are building a technology that will be able to read and understand technical drawings to a level far superior than a human being. It will not only be able to detect changes and potential, you know, big, big impactful mistakes, but also understands how to make them right, how to correct them, to provide you real time design solutions. And the goal is to generate flawless architecture. So I'm imagining the near future, almost like when Tony Stark's seat next to Jarvis, right, and show me all the deviations between architecture and the structural, or we have 250 square feet that we need to add to this. Can you show me different options on the development of the building? It's like, well, let's sum up the conflicts. There's a mechanical conflict in here. Okay, how can we reroute this? Show me two options of rerouting this. Can you see that? And then it starts getting better and better, but you are in control of it. You are, you know, directing what are these potential solutions. Show me what's the most cost effective or less energy consumption solution it is.
Jean Pierre Trou: Data driven design solutions at the architect's fingertips and the engineer's fingertips and at the builders fingertips. Imagine just value engineering of the future won't be, it won't exist because you already made all the smart decisions until the point of construction. And that's what I see for Mbue one day.
Ron Green: So again, it's really augmentative. It's not replacing you. No.
Jean Pierre Trou: not at all. I would not replace myself. So yeah and that's that's what gets me excited and see all these things noise and all these startups popping up and trying to to solve real problems in what makes me so excited about it.
Ron Green: A lot of people are worried about AI potentially replacing humans and taking their jobs. It doesn't sound like anybody would complain about this change control review process being taken away, right? It's kind of monotonous. But are there parts of the architecture space that you think AI won't play and that should be reserved for humans alone?
Jean Pierre Trou: Look, I mean, I was the biggest, you know, my wife works at Kung Fu AI, managing director, I'm sorry, marketing director at Kung Fu AI. She'll be the manager director. She'll be talking. But I always, Tisor, when she started working was, you're creating, you know, the terminator, you know, it's going to take us all.
Jean Pierre Trou: And now I'm funding AI company, you know, in the industry. I see the amazing, you know, the amazing opportunity of AI to automate very tedious tasks in our industry. There are areas in our industry that I feel is where humans provide the most value, which is the design process. And the design process itself is actually flow. Like imagine before when you were writing a paper, right? And there's that concept of the blank page effect, right? You have to get all that creativity to put into words, right? It's very similar in design, similar in music, you know? So having ways for us to incentivize and facilitate and make the decision of the design process faster is very exciting to me, right? So I was very early user of Pinterest, for example, when it started. You know, everyone, you know, it's full of hats now. So it kind of lost its luster. But what I was going with that is that I used to do pin boards. All my inspiration come with my memories and things that I saw in magazines, pictures of magazines. Pinterest game would give me a whole plethora of images that I can use and create a mood board. The same thing that I was doing manually. And now AI can generate different from prompts of my design concept that I control can generate multiple iterations of that. So I already have a very instant mood board that I can curate and get better and better. So I see AI as more of a companion on the design process to it and not to replace the architect.
Ron Green: So more augmentative.
Jean Pierre Trou: Absolutely. I meant that's a perfect word because look, the architect today spends 30 to 40% on activities that he is not the best at, including quality control, right? We all make mistakes. I make mistakes and still those mistakes have a significant impact on the project. We haven't talked about that, but changes what I'm focusing on equates to 5 to 10% of all of all the construction costs in our industry. We built over $1 .4 trillion each year. 6% of our annual GDP is construction. So 5 to 10% of that, we're talking billions, over $70 billion is wasted because we make mistakes, right? And that's what I'm trying to solve. That is what I'm trying to mitigate. And AI is going to be able to eliminate hopefully a large percent of those mistakes pretty soon.
Ron Green: So starting an AI company at this time period, what's that like as a tech entrepreneur? What was the experience? What were the surprises that you ran across this year and maybe some of the hardships?
Jean Pierre Trou: Yeah, 2023 was a very weird year for everyone who started at startup. But we're very happy that we were able to get some funding and get going. It was hard. But what I saw was everyone that I was talking to, I thought that there was this, um, can adversity about AI in our industry. And it was quite the opposite. I encountered this, not the utmost excitement and everyone saying, yes, and you can do this. And they're like, um, probably, you know, I overpromise a lot of things that I have to be able to deliver, but there is such an excitement today about AI and be able to optimize all those studio stats in construction and architecture. The most challenging thing has been, um, you know, with the release of chat GPT, uh, during, um, early this year, all the craze about the AI, there was this illusion that now AI is very easy to make, right? And try to convince stakeholders, investors, uh, that what we're building is not a chat GPT for eggs, uh, which is actually much more complex than that has been a challenge. And I think now, um, our latest conversations that we've been having, people have realized, um, you know, and differentiate all the different areas of AI as you know, you know, it's not just large language models. What we're building is really a multi -model, um, uh, approach, you know,
Ron Green: We're seeing kind of the same thing where initially chat GPT I think in a interesting way got the general public excited about AI in a way they had been before but maybe I think a lot of people over indexed on it and became maybe thought that was the only way AI could be approached or the only capabilities.
Jean Pierre Trou: I mean, it's understandable. I mean, the tools of the large language models in general, I mean, we've seen an increase on their capabilities in such a short period of time. That gives that illusion, it's like, well, now you can solve every single problem in the world, right? Well, try to make it understand that Simple Square is a wall or if a table becomes such a trivial thing, it's something very complex.
Ron Green: Yeah, absolutely. OK, well, this has been a fantastic discussion. We like to wrap up here with just a kind of a like question, which is if you could pick any AI to solve a problem in your own personal life.
Jean Pierre Trou: Oh, my personal life?
Ron Green: Anything you want, what it would be.
Jean Pierre Trou: Okay, I have a good one. My wife says that I'm very positive and I always over commit. And I always say, yes, yes, I'll do it. So have an AI assistant that tells me, jump here, you already over committed for today and for tomorrow and the rest of the week. I think that would be fantastic. So yeah.
Ron Green: That's awesome. I love it. Well, John Pierre, thank you for joining us today. It was fascinating. It's been awesome.
Jean Pierre Trou: Thank you for having me!
