Applying AI To The Construction Industry At Buildots

Hello, and welcome to the Machine Learning Podcast, the podcast about going from idea to delivery with machine learning.

Your host is Tobias Macy. And today, I'm interviewing Ori Silverberg about applications of AI for optimizing building construction at Buildoz. So, Ori, can you start by introducing yourself?

Yes. Of course. Thank you, Tobias.

So my name is Ori. I'm the VP of engineering at Buildoz. This is a company where we use computer vision and AI to transform how construction projects are managed. And my role spans from shaping the technical road map all the way to ensuring our engineering teams have the, you know, the clarity, the autonomy, and all the tools they need to solve the real world construction problems. I was, still am, the employee in the company, and it was

something I did all the way through the past seven years, in Buildoz was to help also shape the product itself and not only shape, like, the way the engine works, but also what the engine does. So it's been a pleasure to work in Buildoz.

And do you remember how you got started working in ML and AI?

So I've always been drawn to solving complex problems.

I love breaking them apart, finding the pattern, finding the structure, and then building them all up with a solution in mind. And One of the things that you learn

is that you can only do it one at a time, if you don't use technology to do so. So my brain is is great and all, but what I really wanted to do was to solve more problems and to do it whether I'm asleep or awake or talk to my, beautiful baby girl. So what I did was I wanted to start using technology, and I created more brains that would work the problem. And the natural next step would be machine learning because machine learning allows me to now not only solve more problems that I could solve myself, but by nurturing the environments of high quality data

and designing effective educational paths, I can actually create brains or, like, a server that can now solve problems that even I couldn't solve myself. This was, like, just, like, a very natural,

development of my my desire to solve these problems. And I think this is what we do in business as well, by the way, is we just solve problems.

You can talk about which, you know, programming language we use, which

type of, which branch of AI we use. But at the end of the day, we have very complex problems to solve. We try and solve them. And AI is, like, a really great tool in your arsenal to solve it.

And so digging into Buildot, so I'm wondering if you can give a bit of a overview about what it is and some of the story behind it and the core problem that you're trying to solve.

Yeah. Honestly.

So the construction industry is huge,

depending on how you measure it, but it's, like, one of the biggest industries,

in the world. And then the, you know, like, the

estimates are talking about, like, 13,000,000,000,000,

of an industry of an industry. Sorry. And

it is believed

to have about $1,000,000,000,000

caught $1,000,000,000,000

of loss caused by delays in this industry. So this industry is not it's very big, but it's not very

effective, and it's not reaching its full potential. So

our founders,

the founders of Build Aft. Right?

They recognized

that despite its size, construction is still lacking something. It's still lacking the real time feedback loops that that allows other industries,

like manufacturing,

to be as effective

as possible, like, to reach their full potential.

And the the idea was

that

we will we will take

we'll take images explaining the reality of what happened on-site.

We'll take the plans.

We'll combine the two. And by creating the way to compare the plan versus the reality and do it often and time after time, we can really speed up their construction and execution without adding any friction to the site crews.

And what we've done so far is that we've analyzed

something that is roughly

three times the size of Manhattan, providing

unprecedented

visibility

into what happens in construction sites,

and we we already help keeping projects on track and on budget.

As you mentioned, construction itself is a very large industry. There's a high degree of variability in terms of the scale scale and complexity of the projects involved. Everything

from I just want to

fix up the creaky front steps on my entryway to I need to build and plan an industrial complex or build a data center or build a skyscraper.

And I'm wondering what axes of scale or what types of complexity you're looking to solve with BuildBots, whether you're focused on residential,

commercial, industrial,

sort of the the categories of construction that you're involved with.

So the project we did was a large scale residential,

project. And by then, I was very naive. I was, you know, at the beginning of a start up, and I really thought that we'll just, you know, keep it residential, keep it, like, small, and then, progress as as we grow. But what we really found was we don't have to do that. We built an engine that can just

digest and analyze any construction

site. And we, you know, we need them to be complex in order for buildups to actually give value. So we would never want to go to, you know, just fixing

the the front door of, of your house. We wanna take the complex project. We wanna take project where

giving this extra edge of, like, optimizing them can really give value to our customers, but we don't limit ourselves to any type of project. So we've so far did large scale residential, large scale commercial projects. We did hospitals.

We did schools. We did anything. We can't even imagine museums,

train stations.

We did airports. We did data centers, battery plants. Even semiconductor fabs, we've done a few of those.

If it is modeled, if it's modeled in three d, we can do it.

In terms of the

challenges that exist on the job sites for these different construction projects, I'm wondering what you see as some of the main sources of inefficiencies

and delays and some of the ways that you're trying to address those complexities

in the process of actually planning and executing a building project?

At it seems like everything is very because, like, you know, no project is anytime similar to another project that we've already did. And even for the construction experts that we work with, every project is special, and every problem and inefficiency

sounds special and unique,

at glance. But then once you have the perspective that Buildlets now has, which is

we see so many projects. We see hundreds of projects of all types. You see that the inefficiencies

just boil down to additive issues.

You'll have different causes, but they all went under the same ones.

And

but it will go, like, one step back. Construction gets delayed for all sorts of reasons. Construction is so complex. It's a real world problem, but, like, the most it's the realest

problem you can see the room. And, you know, external factors like weather, like the scope and design changing, like model updates, the developer wanting to add the floors, to remove floors, to change anything,

it can all all happen. But our data shows that about 50%

of the delays are truly caused by operational

inefficiencies

by people not making the right decision in the right time and not doing it early enough. And

and it's impossible to do it without billouts, and then we'll go into that a bit a bit later on. But what we've seen is that the complexity and the fragmentation of the work just

cannot be cannot be managed, using, like, manual,

manual processes,

as they are without build hooks. So delays, either labor,

miscommunication,

incorrect assumptions,

they're not an exception. They're not something that happens rarely. They happen all the time. And even with project that use build outs, they keep on happening. But with build outs, we're able to nip them in the bud much, much quicker. So but I will give a few anecdotes

because the world

is is, you know, it's full of them. So I won't give any names. But in one of the project, GCT pushed the electricity trade to bring more people on-site because according to the plan, they're supposed to. And each electricity worker will just, you know, install, like, 10 sockets every day or whatever. And another trade didn't complete building the walls it was supposed to according to the plan. So when the electricity trade would have brought more people, they would have nothing to do. They would not have any

available sockets to install. So what happened there was they kept pushing the team

kept pushing the wrong trade. They kept pushing them to do something that would just lead to idle workers, just lead to workers that we pay, that, you know, take

space that we do not progress in any other activities because they are there, they would have been idle. And one other thing that can happen, and, again, I'm like, I'm fixing the numbers so it's easier to understand, but they're not far away from truth, is

say that an activity is supposed to take roughly hundred days. After ten days, the superintendent,

what he does without build outs, what they do is they ask the train informally, are they progressing according to plan? And the suit the train will tell them, yeah. Probably.

And based on that, they will, you know, continue planning. They will continue bringing resources to site, both workers and materials. But once you have correct data and accurate data that comes frequently, you know that in reality,

we're supposed to finish 10%. They can do it without build up. We're supposed to they plan to complete 10% of the work, but actually completed 5%. So now you need to ask yourself, are we do we have just a 5% gap, or are we doing going at half the speed that we need to? And when you have frequent data captures, when you have the correct data, you know that they're going at half the speed. And then what you know, only after ten days of starting the work, you know that they're going to be a hundred days late. This changes the picture. This changes the importance you put to this issue and how many focus and how many workers you wanna bring on-site. So with build outs, this is just what happens. You find the issues early and you solve them.

For projects that aren't relying on build outs, obviously, there are numerous construction is an industry that is as old as civilization since we started building shelter. And there have been many different ways to address some of the growing complexity of these construction projects, particularly when you're talking about industrial projects, things like skyscrapers that have a lot of interleaved requirements and very specific sequences at which the tasks have to be completed. When you're working with

potential customers

to onboard them,

what are some of the typical ways that you are seeing in terms of manual and technical processes that the industry has

typically relied on to be able to manage these sources of complexity and delay and

waste? So managing a construction site is something that used to happen. There'll be lots of, unfortunately, for them in the all the that are not using pillar so far.

It happens,

with a lot of manual processes. So there are a lot of people walking the floor with clipboards, filling out spreadsheets, marking it down on their three d maps. So the execution itself revolves around a lot of manual processes. By the way, their processes are amazing. They this industry works with really low margins, and they have to be excellent to the best of their abilities. They have to be great. They have to plan well. So or you want planning,

you continue on that. Construction is like building a waterfall project. Right? You know the end result. You have to build it all the way through, and you have to do it efficiently,

as efficiently as possible.

So the industry really evolved,

their planning capabilities.

So based on their lack of information they have, they built the best processes around.

And

so they built this idea. They have, legal construction. They have, like, this sort of, agile methodologies.

They replan every week. They try and figure out what is the best way to move forward, but they're doing it all manually. So they use, like, sort of like the old agile

way. They have sticky

notes pasted on whiteboards

saying what they should do next. They have people walking around saying, I counted

100 sockets installed out of 120

that are supposed to be installed on on the Second Floor, but it's all with it's all very manual, and it's all very based on gut feeling

of the construction expert that worksite.

I would say it's not it's not their fault. It's not their fault that they can that they sometimes even just ignore even the manual tracking. Just doing it is is most of the time isn't feasible. You cannot manually track hundreds of thousands of of elements of small tasks and, like, figure out what are the final updated data results.

And

the

information age, the Internet, these have had massive impacts on many areas of industry

and

ways that we work and even introducing new categories of products. But with construction

being firmly cemented in the real world, it seems as though it's very resistant to digitization.

Obviously, there are project planning tools that you can bring in. There

are computer assisted drafting that has accelerated the accuracy

of being able to build blueprints and transmit those building plans. But I'm wondering what are some of the ways that the construction industry is resistant to digital transformation

because of its very physical nature?

of all, go back to my prototype,

to my prototype answer. Construction is

fundamentally a prototype industry. No two projects are exactly the same or behave exactly the same. And each site is is in has its own operational environment that is being actively built, throughout the prod the project itself. Often without reliable infrastructure, they don't have electricity.

To be in Florida, they don't have, they don't have Internet. Everything is packed with heavy equipment. There are logistic challenges. There

there are different trades. All of them come together for for a single project and then disband to move on to to the next ones.

And many of the workers are just are not tech savvy. And the construction workers themselves, you can ask them all you want, like, to fill out forms or to use tech, but they they may not know how. And everything here is just like a construction project. It starts.

It sort of build its own ecosystem and then expands afterwards and starts again somewhere else. And in this environment, it's very hard to streamline anything new. It's very difficult to change anything. Whatever is, like, the norm in the industry

sort of keeps being the norm because you you keep getting new people. You can't change their behaviors or teach them,

new technologies very easily. But I would say that I think that a lot of things have changed in the construction industry. Planning

moved through

amazing,

changes and revolutions

in the in the past years.

The industry moved from working with two d drawings to having,

important, like, three d modeling that is super precise and exact, and we use that. And the fact that the environment itself has a lot of new tools, a lot of planning tools, a lot of management tools, really help build that

reach a place where we can actually build something that uses all of this information that is now structured and in the cloud.

So digging into some of the technical applications that you're building at Buildoz, I'm wondering if you can talk through some of the ways that you're able to leverage AI and machine learning to this very complex and messy real world problem.

Okay. So BuildOps AI is not, like, sprinkled on top. This is the core of what we do. BuildLA's could not have been created ten years ago.

AI had to reach a level of maturity,

that it that it had in recent years that made it possible to finally tackle

this real world problem,

that has such vast complexity

that is, like,

fixing or optimizing construction sites at scale. So what we do is with a computer vision to extract

data from noisy, imperfect three three hundred and sixty degrees side visuals. We take three d models of millions of elements

and classify them and, like, extract.

And we have to we have to separate between cold water pipes to hot water pipes. And all we have are these,

like, unstructured metadata

field that are inconsistent

all over the world. Like, no architects in The US or in The UK don't use the same the same attributes on these elements. So we have to classify three d models. We have to detect,

precise locations and to

unblurry and,

and to white balance the image that's coming from these noisy construction sites, where everything is always a bit darker and a bit blurrier than we would hope and compare it to. And what we do is we actually build like this engine that creates layer,

on top of another layer

that always enriches the data to get to the next step where we can run more algorithms.

Some of them computer vision, some of them CNNs,

some of them are just, like, data models to get more and more highly enriched insights. So we start from, like, images, and we get all the way to, like, predicting,

what is going to be the delay of

whatever of, painting the walls, on level five. So getting from here to there means we have to use an arsenal of a lot of algorithms.

No one algorithm is going to be to be enough.

And

the other aspect of bringing AI to bear on this problem is that you need some sort of digitization

to act as inputs to the AI to be able to operate effectively. So, obviously,

in the very early stages of the project, a lot of the

inputs are already going to be digitized,

the three d models, the CAD designs,

the bills of lading, the inventory,

etcetera.

But as you actually start moving from that planning phase to manifesting things in the real world, that's where things get messy and complicated. And so then you have to be able to

transform

those real world on the job realities into more inputs for build outs to be able to operate on. You mentioned computer vision, but I'm wondering if you could talk through the elements of data collection,

particularly ways that you're able to automate it so that you're not adding yet another concern for the people who are already very busy trying to get the job done.

You mentioned the three d models and all the already digitized inputs. And I wanna say that even there, there are a lot of complexities

because, as I said, three d models are not all created according to some, you know, very strict protocol and are built in the same way. And what we find is that construction sites have such tight date such tight date lines that even when we begin even when the site actually begins construction, not everything is already modeled. And they do it, like, as they progress progress as they go, and they update their models, and they start with design model and then, you know, fix everything that they need fixing throughout the life cycle of a project. So even there, we have a lot of of complexities and then and missing parts and that we have to somehow solve to give a reasonable product. But when coming to the visual and and the reality capturing part of it,

So the input that we started with was 360

degree cameras.

And the reason for that is that we wanted to,

disrupt the site

team as little as possible. And the three hundred three hundred and sixty degree camera, they can just walk into your room and walk out. You don't need to do anything else.

So

today takes

one or two minutes to just walk a relatively large floor doing a doing a dedicated

dedicated capture.

But from there, we evolved

we evolved to not rely solely on the 360

degree, cameras,

because when working outside, when working on vast spaces,

they're just not good enough. We work on project that span

many square kilometers.

And people who would walk there, would just, you know, have to

will have to open their their tent and spend the night. So what we do there is we use drones.

And I think one of the most amazing thing that I've seen our infrastructure,

we've one of the most amazing things that I've seen our engine

adapt to is actually changing the input

from these, 360

degrees, cameras into drone footage

with just, like, a few weeks of work. Nothing more because just we just build the right engine. We build the right infrastructure.

And, by the way, this is something that we constantly work on, like, making sure that we're not tied to any sort of input, to any sort of specific construction site. We want to be able to adapt quickly to the change of ways of of construction. And so another thing that we did about a year ago, we introduced

lidar scans, into build lots,

which also felt like like it's a great achievement for our infrastructure and design that just adding another input is just as easy as plug and play.

So for construction crews that are using BuildBots, I'm wondering if you can talk through some of the ways that it shifts the overall approach to the project and the ways that it integrates

into their work day to day and from start to finish.

Until recently, we focused on helping the GC site teams and owners,

deliver projects more efficiently.

Delays, they are more on budget, etcetera. It was by spotting activities at risk of delay. It was it was by resolving unfinished work, simplifying reporting, or

just simply improving the the collaboration on-site.

But we've now significantly

widened the scope. We realized that by having

this sort of data, we can change many more processes

in the life cycle of a GC company, not only the GC site. So

now in Buildoz,

Buildoz is now not being used only by the site team, but also by senior executives

to manage the entire portfolio. So now not only

point you as to the biggest risk in a project,

but which projects have the biggest risks,

which projects need more focus,

from the, from the senior executives teams as well. And now they now build outs can be used much earlier in project life cycle. And

even before, by the way, the bulldozer reaches the site. When you know how construction works, when you have the

benchmarks

of how things can progress,

even looking at a schedule and three d drawing

without even seeing the actual footage from the site yet, we can already tell you if things just don't have a good chance of working as planned. And if what your planning is just unrealistic, we need to, revisit your plans from the beginning.

So I would say we're working with an industry that created and follows a lot of great processes and the processes of trying to handle the lack of transparency and predictability.

And what you get when you do when you introduce into these already existing processes

correct data, when you input the when you input the right information, you do it at the right time, and you do it in the right format, all of the existing processes just start making much more sense, and the outputs create planning and resource allocations. I wanna give another anecdote that is relevant, like, to to this sort of, like, how to current processes simply change with BuildOps. There was this one hospital project in The UK, where the superintendent

noticed using Buildbox, that some delays starts to accumulate

for the electricity trade. And they asked the trade, how can it be? Since they knew that there were many electricity workers on-site. And soon they realized,

by talking to the trade and looking at the data in Buildbox, that the electricians waited for another train to finished. And there they waited for them to finish their raised access for installation. And they waited because they knew that there is this this dependency, that they couldn't continue working until they finished their raised access for installation.

But it wasn't really needed. It was something they decided they wanted to do, back when they started planning the site. But in reality, they just could eliminate this decision. They could reverse this dependency,

and it sped up the progress of the project

immediately by with a blink of an eye. And this thing could only happen if getting data was so easy. By having this discussion with the electricity trade, they could just pull with it whichever data point they wanted.

And so digging now into the

engineering work that you've done to create build ops, I'm wondering if you can talk through the technical architecture and design of the system that's powering this experience.

Okay. So I'll do my best to find that without needing, like, a football field sized whiteboard

behind me.

Okay. So I'll look at what our technology is doing as, like, generating data, generating,

more

data that is getting more enriched and more complex the more you continue and and, like, getting away from the source, which are the three d models and the videos.

So we create a detailed digital twin, a digital representation

of the construction site in our database.

Doing that is like taking a real world problem, something that, you know, people sit in university

learning, like, all the complexities, all the dependencies, everything that explains what a construction site is and modeling it

in, you know, in a database

is complex. It's complex because you have to now enforce relationships. You have to enforce what it means for an activity to be. Does it always belong to a single trade? Is a level always in a single building? You have to model everything and to create the restrictions,

and you want to limit the flexibilities,

if you model as much as possible

so you have easier time creating the next features. And when you build the right model, and we believe that today we have the right model or at least a right model, this allows our construction expert experts to now onboard any site, any construction site, a hospital, a school, as I said before, a residential, commercial, and deciding to build up quickly and accurately.

And what we do afterwards is we whenever new videos are uploaded,

when a construction site somewhere in the world uploads videos into build up, our Kubernetes clusters,

spin up thousands of cloud servers running dozens of computer vision and deep learning algorithms in parallel. And this is where the raw imagery gets turned into

structured,

actionable data. And finally, once we have all that and we have the plans of what is supposed to happen, we have the reality of what is happening, as captured, in the images. What we have done is finally,

we can now

take all these new data points. Right? Finally,

every time a new data point is added to the platform,

from a processed video, from an updated model, or a client schedule change, our engine can now reprocess

everything. They can reprocess the full context of what is happening on-site, update the progress tracking, to update the deviation allowances, to update the risk forecasting, everything.

And I think that one of the most interesting things that happen in build outs is that

the reality

changes often. So many times, what we'll have is, like, a a client,

telling us, finally, I have this three d model of the, whatever, the MEP,

ready for you to analyze. Can you please analyze back, like, the past four months and imagine as if this model has all had always been there? And we had to build an engine, a platform that supports that, that supports any change, supports any update by the user, and to fix the history, to fix the present so we can fix the future. Had to build that, something very robust.

Working with construction projects, dealing with

real world requirements,

dealing with the creation

of structures that need to be safety, need to conform to various regulations.

They need to be able to withstand whatever environmental factors are present on-site.

That brings a lot of risk along with it. And so given the fact that you are providing

inputs to the ongoing work to actually create these structures that I'm sure infers a certain amount of risk to you as well.

So given the safety critical nature, the regulatory nature of the construction industry, I'm wondering how that influences

the ways that you think about which types of AI models to use, the selection of those models, and how they actually get applied, where to

industries,

which I'm sure, in many industries, you have to do that as well. But we also prioritize

explainability.

So our models

are not designed

to replace

anyone. They're not designed to replace the decision making process. They're designed to assist the decision makers.

And to do that,

we provide clear,

auditable data trails and not black box outputs.

So

every data point you see on the Billups platform, like any user, any data point you see, you can track back

to why this number is what it is. And if you see that buildouts tells you that you're going to be late unless you change anything for this this reason, you can now dive into all the elements that we've seen, and you can see the imagery of, like, this socket was installed. This socket was not installed. This one was installed,

two weeks delayed.

You could see all of that, and always, it helps us by building trust. It helps our clients by having the proof to show the trade with the subcontractors

and everyone else they want to, they want to to actually

explain why they're making the decision that they're making, and it helps them by verifying that what we tell them is indeed true.

Generative

AI and large language models have

brought in a new style of user interface to numerous different products. I'm wondering how you're thinking about the application of AI to

simplifying the ease of use, the onboarding,

the styles of interaction

that your customers are having with BuildOps so that they don't have to necessarily spend a lot of time figuring out what are all the different knobs and dials and nooks and crannies of the product. I just wanna know what do I need to know and how do I move forward.

Okay. So I'll start by saying that we are still learning. I think that GenAI is like, you know, taking the world in a storm, and and

we still experiment with how GenAI can help us simply be faster and better at what we do. You know, just building the right platform, but also how to use it, like, more exposed to the clients to help them get more from below. So

two things that I'll say that we already have, facing the customers themselves.

You know, they'll take three. So the would be we introduced

a few months ago, we introduced

chat agents,

an LLM based chat agent that

exposes the build up's data,

or actually the user's data,

as I generated the build ups to the customers.

So now customers can instead of understanding how to use, you know, our,

pretty great UX, but instead of using that, they have this specific question in mind that aren't in a training. They just wanna know something. They can now go to this chat agent, just ask them wherever they wanna they wanna ask. And what happens behind the scenes is that we understand what are the entities they're referring to, and

the LLM generates an SQL query

that queries our database

and outputs the results and delivers it to the client. And

this allows our customers who are still not,

fully worried on the platform

or just don't have the time to now, you know, traverse their various screens

to get the input they need as quickly as possible.

So this is, like, to the customers themselves.

Another thing that we're doing is we're now creating amazing onboarding videos

using Gen AI. So we have, build those avatars

that are now,

the face of build also, like, extending every new feature. And now onboarding clients can be much more streamlined using very fine tune, videos. So this is the And I would say the is Buildoz is building, at the end of the day, a b to b software. And we're building something that helps our customers

improve and optimize the work by so much,

that since Buildoz is a b to b,

a b to b software company,

we do not shy away from building complex

software as long as it produces enough value to the customers,

we need to earn their attention.

The AI landscape

is obviously under constant evolution. It's moving very rapidly, particularly in the generative space.

And

I'm wondering how you're thinking about the

future trajectory

of build outs and the capabilities that you can add in light of the those new underlying

model capabilities,

new architectures,

new

ways

of structuring

the different models, particularly thinking towards agentic use cases and just some of the ways that you're thinking about the potential for

bringing even more

capability

to construction crews that are relying on build outs for their project management to be able to get more useful insights, maybe

even help with the generation

of

blueprints,

construction plans, etcetera?

Yes. So I think that

GenAI

is very useful

in

helping people interact with computers,

but you need to have, like, reliable

data behind the computer in order to to streamline and actually get value if you don't want to just, you know, help the user refine your own thoughts. I think that Jenny and I may give build outs, like,

another step, another

very tall stack to stand on. We now work on

helping,

helping customers generate plans, generate,

their their program, and we build lots now know

what programs are possible, what programs are good, and will lead to their maximum efficiency.

And the ability to now have this,

this feedback cycle of a customer, like, just telling the the computer what he wants, what he wants,

can make this process

much better. It can now take the data that Google has and actually

take all the needs and the requirements that a user has that we will never be able to encapsulate.

Only this type of gen AI can help us do it very efficiently.

And I think this is, like, an amazing,

an amazing thing that we could do in the future. But as I said again as I said before, I think that we're still learning.

One of the things that we are doing inside the company and that we're taking each department,

sales, marketing, engineering, algorithm,

research, and we make sure that they are all

trying to

improve as much as possible by using GenAI. So you have all these Slack channels where everyone who notices anything anything valuable that they've used AI for, and they post it there. So everyone can now try it. And we use,

GenAI based IDEs,

and we use, Gen AI when we write emails and when we when we draft messages because we want to hone our skills and to constantly understand what Gen AI can do for us.

And from there,

we believe we'll get more and more ideas on what Gen AI can do for our customers as well.

Another element of the construction industry

is obviously all of the paperwork involved in getting building permits,

verifying that you're in compliance with the local regulations,

but also the

systems within which the projects exist. So things like the electrical grid, the effects on the local environment,

whether you're disrupting any

natural water sources, etcetera, etcetera. And I'm wondering how you're thinking about the potential

for build ups to be able to assist with,

some of that tedious work of the paperwork as well as some of the

planning around the broader system in which the project exists.

of all, I'm not sure that I ever thought of that.

So I'll try my best. I think that in general, in the construction industry, what we are seeing today is that more and more,

startup companies and even old tech

are finding their way around this industry. We added more and more data sources,

in more and more parts of the construction

process,

from before a project site even exists, from before

the bidding on an on the future construction site even ends.

More and more data is getting digitized.

And I think what we'll see is that the,

integration

of more sources coming together is actually going to bring us the next leaps. And I think what Builders brings is

the data of what truly happens in construction sites. I'm not sure how it will,

benefit the the domain that you just said, which is you just mentioned, but I'm pretty certain that only by integrating

all the data points, you can get to the true,

optimization of of this industry.

Yeah. I I can also imagine some potential use cases on things like urban planning where you say, okay, I know that this construction site is happening over here. This other construction site is happening over here. How is that gonna impact traffic patterns,

electrical loads, loads on the water supply, etcetera, etcetera.

Understand. Okay. So

one of the things that we are creating is the ability to classify and analyze the three d models

in a way that actually,

correlates to the to the design of the project, the way the project is going to be managed,

and

and actually be built.

And

from there, we already see many insights that we can bring to the table, like possible clash detections, like, where,

misplacing a pipe by even just, you know, a little bit is going to create havoc because you won't be able to install another element.

And I think when we'll be able

to take larger models, not only, describing the construction site themselves,

we may be able to get, to what you're saying and, like, you know, getting it all together.

And in your work

of creating the technology behind Buildoz,

working with the customers,

working with the elements of the industry, I'm wondering what are some of the most interesting or innovative or unexpected ways that you've seen your product

used?

One of the most inspiring moments, at least for me, came from a project manager

in one of our early UK projects. I'm from someone we now consider a true super user and a true champion,

to BuildLAX.

And after walking with Buildouts once, he refused

to ever join another project that doesn't have it. He had an incredible ability to turn more data

into into insights. So we have this feature that just

exposes

a lot of just like a gigantic data table on the project. You can see anything happening for each area, for each activity, like a very, very massive table. It's there. It's there for, you know, just diving into the tiny bit of details. And what he did, he was somehow able to identify trends, to identify the risks, the one that even we didn't envision. We didn't

figure out could be extracted from the data yet.

And his creative use of the system actually inspired us to develop new features

that made those kinds of insights now accessible to all users, with his permission, of course. And this was a probable reminder that the best innovations

often come from the people closest to the problem and not from, you know, people sitting in their

cushion,

cushion that sits,

in the in the store company far away from them.

And in your work of building this technology, learning more about this particular industry and the challenges that it presents, what are some of the most interesting or unexpected or challenging lessons that you've learned personally?

So I think that one of the biggest lessons that I've learned, it's very counterintuitive

to someone coming from the tech industry, is that clarity

almost always beats complexity.

It's better to be simpler than being more precise and more accurate.

Even the most sophisticated algorithm, they just don't matter if they don't produce simple, actionable insights.

And customers will happily trust and act on a data point they can understand and validate

even if it is less

accurate, but they will hesitate if it feels like a black box. So this is one. Another interesting engineering challenge we had is managing our assumptions

on how construction site behave. So as I said before, we are modeling a construction site. When you do that, you you need to make assumptions on how things are built and what

every construction site in the world has to to buy back. So let's take buildings, for example, very simple very simple example.

Let's take building. When you talk about a level,

you could assume that a level will always reside in a building,

But then you get to the next construction site where there are two buildings,

connected by this, like, bridge floor.

And now you scratch your head and wonder, where is it? Do I need to now remove my assumption that the level always reside in a building? And if so, why it's going to happen there? It's like many to many relationship, and how do I not change my entire practice for that? So just ignore it. Just split this floor, have to build a a, have to build a b, just move on with my life and, like, hope it will never come back again to haunt me. So modeling things well is very difficult. Modeling a real world problem by construction is

honestly possible, but you don't need to do the best job. You just have to to do a good enough job. You have to model the world. It's in a way that is simple enough to actually use and build a platform on, but complex enough to support the value that you want to provide to the customer.

And as you continue

to build and iterate on the product, you continue to explore and stay apprised with the AI industry that you're

operating within. What are some of the things you have planned for the near to medium term of either the build outs product or the potential applications of AI to this problem domain?

Yeah.

So we aim to make construction companies

more performance driven. We want to help them achieve operational excellence, and we want to help them deliver projects on time and on budget and

and everything else that our marketing department,

can can articulate so well. And to do so, we need to give the construction team superpowers. We need to give them the ability to see risks before they materialize

and need to mitigate it to help them to mitigate delays before they escalate.

And we have to be able to predict better with less data. We have to now take not only the data that is created in the current construction site, but to learn from all the construction sites that we already did.

So we can use, like, the one, two, three data points in the construction project to already generate,

the insight that will have this team,

optimize their own car building. And to do so, we have to be much more accurate. We have to always push forward with, like, modeling the world better to find all the dependencies and to find all the connections

that we now miss. And this will result in creating much more complex algorithms that require

a lot more data, to actually we can actually build. Something else that we have to do is to expand into more phases of the project life cycle,

into superstructure

and, and the facade and everything else that revolves around, like, the building itself that would be, like, the full picture.

Are there any other aspects of the work that you're doing at Buildoz,

the challenges

of providing

automation

and insights to the construction industry

or the overall potential for applications in AI to construction and real world problems that we didn't discuss yet that you'd like to cover before we close out the show.

Good. So I wanna add something that is, like, the way we tackle problems, build outs, and the way we see AI. So to us, AI is is one of the most important tools that we have, but what we are trying to actually do is to solve problems. And I'm coming back to the beginning of the show, like, what drove me into machine learning. So at Builders, we're trying to solve a problem. Our problem is to how to optimize the construction site, how to deliver these insights. And the way we do that is by taking a problem and always

create

engineering and algorithms

and an AI that helps us split this problem into smaller problems and to keep doing so until the problems are just small enough for AI to completely automate them. Then what you find in a real world problem is that it's never easy. It never is like a very simple problem, a pretty fine problem that you just know in the beginning that you just need more data to solve. What you truly need is to uncover

the true problems that lie underneath the big one, The what I need to solve in order to reach this final final solution is a big problem. And I think this is what we've done for seven years in Build A Lot

is to continue

take problems apart, to continue spreading them more and more until they are so easy to explain to to another human being and so easy to explain to a computer,

and to a CNN, to to an algorithm

on how to solve. And only when you reach this granularity of problems, you can actually move on to the next challenge. And I think this is the way we see AI. AI

helps

us split the problem, and AI helps us solve the the smaller problems. But we still need, like, the people,

behind it. We still need the creative people to see the patterns, to use them all, just to see the patterns now and to understand how to truly solve the complex problems.

Alright. Well, for anybody who wants to get in touch with you and follow along with the work that you and team are doing, I'll have you add your preferred contact information to the show notes. And as the final question, I'd like to get your perspective on what you see as being the biggest gap in the tooling technology or human training that's available for AI systems today.

I think that at least what other than build us is that it's all about the context. You need to give the algorithm

you're building, the LLM or whatever proprietary,

algorithm you're building. You need to give it the right context. You need to understand what data you actually need to fit it, in order for you to get to the right to the right, answer. And to understand what is the true context today

requires creativity. It requires

legwork of understanding

the problem. It requires us,

build outs to go to site teams and to go to construction sites and and understand

understand what it is

to build and what it is to manage, what it is to explain a construction site. And only once we do that, we're able to now, you know, fit it back to our algorithms and fit it back to our engine and our,

in our software architecture so we can truly solve the problem. So AI

is a great tool. It's a mandatory tool if you wanna solve complex problems. But unless you give it the right framing, the right context, the right the right data, not just tons of data, it will just not work.

So to me, I think being able to do that is one of the things that's that stops, like, the

experience and talk to the individuals from, like, solving the next problem.

Alright. Well, thank you very much for taking the time today to join me and share the work that you and your team are doing at Build Dots and just the overall challenges

of trying to

add digitization

and operational efficiency to such a inherently physical real world problem. It's very interesting project, interesting problem area that you're attacking, and so I appreciate you taking the time today to share your thoughts and experiences on that, and I hope you enjoy the rest of your day. Thank you the best. You too.

Thank you for listening, and don't forget to check out our other shows, the Data Engineering Podcast, which covers the latest in modern data management,

and podcast.init,

which covers the Python language, its community, and the innovative ways it is being used. You can visit the site at the machinelearningpodcast.com

to subscribe to the show, sign up for the mailing list, and read the show notes. And if you've learned something or tried out a project from the show, then tell us about it. Email hosts@themachinelearningpodcast.com

with your story. To help other people find the show, please leave a review on Apple Podcasts and tell your friends and coworkers.

AI Engineering Podcast