Introduction

Very nice to see everyone. Thank you, Andreas, for a nice introduction to what I'm going to talk about.

And it was interesting because I came down and asked Andreas, what's the audience profile like? Who am I talking to?

So Andreas said it's diverse, it's random. So let's start with a question.

With a show of hands, how many people are working in technology or technology -adjacent industries anything from like software developer all the way to it and all that okay fair how many

people are working in say services which are not technology or technology adjacent so doctors other service providers okay and yeah and fair and anybody working in construction construction or construction -adjacent industries.

Okay that's perfect.

AI in Construction Compliance: Focus on the Netherlands

So this is going to be a very interesting talk, because I'm going to show you what, like Andreas mentioned, governments, architects, project developers are working with in terms of AI.

And of course, I'm going to focus on one aspect of it, right, there are a dime a dozen tools out there that people use, but I'm going to focus on construction compliance in the Netherlands and what's happening there with AI.

Motivation and Problem Statement

Full disclosure, Struck is a company that I co -founded with my partner Max, and we essentially, our aim is to accelerate construction compliance.

And why construction compliance? Because of this.

Compliance delays are causing delays in construction, and we know that we have a housing crisis here in the Netherlands. And I found that out acutely, sorry, before.

So, like, the delays have been there all around, and the delays have been there for years, right?

So AI has reached sort of mainstream knowledge, say, 2022, 2023, with the advent of ChatGPT, but housing crisis has been something that we've been seeing across Europe, across the world.

And it's not just a housing crisis, right? It's a general infrastructure crisis as well.

More things need to be built, built better, built more sustainably, built in a way that lasts longer and one of the biggest bottlenecks is compliance and if you

Scale of the Compliance Bottleneck

look at some of the numbers here you will see that 1 to 8 years even before you lay a single brick or like even do anything on the ground is spent on compliance on design on reworks over 5 billion euros is lost annually just in the Netherlands due to compliance related issues and of course there's a

lot of regulations right there are there can be two solutions to this we can say we build without any regulations? That doesn't really sound good because we do

want to live in these houses, we don't use these buildings, we don't use these bridges, right? Or we work on ways to simplify compliance.

And that was our vision.

Why Compliance Happens Too Late

The first thing that anybody in this space will tell you is that one of the main reasons a lot of money, time is spent on compliance is because compliance works as an afterthought.

Decisions are made, made, decisions are taken, designs are made, plans are made, and compliance is only majorly addressed when you actually need a permit, when you actually go to your municipality,

when you go to your local government, your provincial government, tell them that you're going to do this, that is when you figure out, oh, wait, I really can't do this at this address, but you've already kind of put money into it, right, you've put time into it. And that's a huge problem.

The Struck Approach

So our idea was why not bring this forward, why not bring this to a stage when those decisions are made. And that was how Struck began.

From Generic AI to Domain-Precise Answers

So what did we do? We start off with something that everybody here has probably tried out, ChatGPT, right?

You kind of ChatGPT, you'd like ask ChatGPT and find out what can I do at this address?

What can I, can I build a bunch of stairs of this width? Can I build a bunch of stairs of this height?

Limits of General LLMs for Compliance

But the problem with ChatGPT, the problem with OpenAI tools is compliance is a very precise, for lack of a better word, science. science, right? Compliance has to deal with rules in a very precise manner.

It is not good enough if you are 0 .1 meter off in the rise of a stair, because that is not compliant anymore, right? It's a fire hazard.

So, at the end of the day, this has to be very precise, and a lot of these data sources are just nowhere to be found.

You might have the most intelligent person in the world, but if the person has no idea something exists, they cannot do anything think about it, right?

And ChatGPT is not the most intelligent person in the world, just to put things in perspective. So this is a problem.

So what we started off with

Making Regulations Accessible and Contextual

was making these regulations accessible in the context of what you're working on, right? So I'll go into a demo very quickly, but just to show you, we started off by making the

building code, what is now called the Bessluit Bouwwerk en Leeuwengeving, part of my Dutch, much, accessible to architects, project developers, and municipality workers. And this soon grew into the largest library of building regulations in the Netherlands.

But this was interesting, because now you would have assumed this was a simple thing people would have thought about it, but no, the problem here is both in terms of access and being precise in terms of how you provide these answers to the users.

Beyond Building Code: Zoning and Local Nuance

Soon we expanded into a lot of other things related to construction, so zoning laws. loss.

And zoning laws are very interesting because they are diverse, right?

A building code is national.

Everywhere in the Netherlands, your rise of a stair has to be x meters, 0 .22, but like x meters.

But it is different rules applicable to this plot of land and a plot of land like 100 meters down the road.

And this is very interesting because these rules Rules are built in ways which have exceptions over exceptions, judgments over judgments, and hierarchy over hierarchy.

It is not very easy to navigate, and I'll show you that in a second.

From Answers to Workflows

We soon expanded to construction compliance across the board, where we provide our users with specific answers regarding questions they might have about their projects, specific to their location, specific to the rules that are applicable to their project in the context of their project.

And after that, we started building, I think two of the talks are gonna be about agents and workflows, so we started building workflows into this because the idea was everybody wants to do this end to end.

You wanna start with a problem, reach a solution, and call it a day because at the end of the day, an architect wants to design, a project developer wants to develop. Compliance is not the most interesting task, right? right?

So what they want is a solution to get to a point where they know what is possible, where they know what is not possible, and try to work within those constraints or negotiate with those when they want to work outside those constraints. And that is what we do.

Core Capability: Verifying Designs Against Regulations

So what we do is the verification of regulations versus your designs. That is struck at its core.

And before I get into the demo, we've been working, like I said, across domains, right? So utilities, with architects, with builders, installers, because compliance touches everyone in the space.

You cannot build something in any part of the value chain if you're not compliant, which is good. Again, rules are not the problem. Too many rules, of course, can be a problem, but

rules are not the problem because we do not want to live in a house that has mold that can fall down on us any point when we're asleep or catch fire, right? So that's not something we want to do.

and in that sense truck tries to streamline compliance with this I'll I'll

Demo: The Old Way vs. The Struck Way

show you what is the current or what used to be the way people worked with these regulations right so let me pull up my window here all right there we go

Manual Navigation of Regulations (Old Way)

so I'll start with a simple thing so this is how it works so I go to BBL online. So BBL is the National Building Code, just to be very precise. So now what am I doing?

I'm going to click on this. I'm going to go here. So I've landed on the BBL.

We'll assume I'm an architect today trying to design an apartment, right? And I need to put in stairs that go from the ground floor all the way to wherever it has to go.

So there are rules about stairs, stairs, and I'm building housing. So how do I find out what those rules are about stairs?

So what I have to do is I have to go here and type Trappen. It's in Dutch. And my Dutch is very narrow and limited to building terms only, but it works for this presentation.

So if you type Trappen, you get Article 3 .21, 4 .27, 3 .20. So you see all the results that that show up, right? But now, you do not really know what you want.

What do you want? You want to find out the rise of a stair. So now, you've got to go through each article and see if it gives you the rise of a stair.

So now, this tells you, this gives you the breadth of a stair, not really the rise. This gives you something else. Again, the breadth of a stair.

And you can go here, I think this should give me the rise of a stair, yes. Yes.

And then again, this is different if you're renovating an existing building versus a new building. How do you find that out?

You've now got to go for ,, which is existing building, right? So if I type this, I'm taken to chapter 5, which is not right. But you will find that chapter 3 is related to ..

So just to show you the utter sort of tediousness of this process right so this is what people are dealing with even people now who do not use

drug are dealing with this the other option is to go to chat gpt and ask a similar question and there you are throwing it's a flip of a coin because chat gpt does not tell you where it is getting this information from and this was way more acute in the beginning of 24 because the

building code was introduced on january 1st 2024 and of course the closed lm models that are out there that you can use do not update their information that rapidly right so

Targeted, Source-Linked Answers (Struck Way)

what does it look like if you are using struck I'm gonna very quickly go here this is gonna be in English this is an example question of course but you can type your own question in there and you can ask what is the maximum rise of a general access there so the interesting thing here is you need to be precise

exercise. You really need to know where are you getting this information from and what is happening.

I really hope this demo works because demos sometimes go awfully wrong. But you really need to know where you're getting this information from.

So the first thing struct tells you is you're looking for this and I'm looking for it in this document which is the building code. So you've got your first level of confidence.

But that's not good enough because you need to know if it has picked out the answer from the right part of the building code.

And as always, demos take longer when there are 30 people watching. We'll give it a second. While it does its thing, let me try to pull something up here.

As a cooking show, I have a pre -canned recipe. So there you go. So while that takes its time to load,

what is the maximum rise of a general access stair? And it will tell you for new buildings, according to table 4 .26, this is what you need. But that is still not good enough.

You need to look at the document because, again, you are responsible for ensuring that it is correct. The AI does not do the job for you, right? At the end of the day, you're responsible.

So there, you click on this button, and then you see that this has been highlighted, and you can find out that this is where it is getting the information from, right? That is talking about reliability.

I see that I'm actually running slower than I expected, So I'll move along a little faster.

So this is what we do with building regulations, right?

You can ask questions about local regulations related to what can you do on, say, Stuyberg Island. Can you put up a duck up bow or like an extension on your roof?

You can ask specific questions, get specific answers, always with the source, always with the source.

Workflow Example: Permit-Free Building Checks

I want to show you one very interesting thing that comes up when we talk about workflows.

Where's my mouse? Here we go.

Context: Permit-Free Building Rules

So in the Netherlands, you are allowed to build on your plot of land without a permit. Most plots of land, like 99 % of them, are allowed to build without a permit. But how much you can build without a permit and where you can do this is limited by a whole host of rules.

Every plot of land which has a house or an office or a main building has a regulation.

Automating Municipal Front-Desk Queries

right so that is what we ended up building for municipalities because this was what happens they get a call at their front desk with people asking them can I do this on my plot it's a very low effort call for the resident but for the municipality they've got to be very precise about this again right so what

do they do so they had to then go to the map look at the plot of land measure it then draw a whole bunch of things and get the answer what we did what did we we do, we kind of put this, we give this AI superpowers, as my colleagues like to say.

Seven-Step Guided Flow with AI Agents

So it's a very simple seven -step workflow, out of which two of them are just confirmations. So you basically say, is this your location? Yes.

Is this a monument? We find that out for you. There's an AI agent that looks whether this address is a monument and tells you it is not or it is.

Is it located in a protected townscape? Again, agents which do this job for you.

And if you looked at the questions, right, all of them are always mentioned with with a source. So you will see that according to the right deans for cultural earth good, you will see that this is mentioned.

No, it's not a monument, so there's no source because it's not a monument. But if it were a monument, it would tell you where you would get it from. 1And everything has to be done with the source.

So you do a few small clicks, you select your roads, you draw the original main building, which the AI already detects for you with this little red square. You just need to kind of trace over it. I'm going to do this very roughly, and we go down.

You indicate the front of your building. If there are any existing buildings, you draw them. Again, the detection has been done for you. You need to confirm it.

Again, we pick up all this information about zoning designations, what is allowed, what is not allowed.

If everything is done correctly, you give it a yes, and we tell you you can add 54 square meters in the pink area.

End-to-End Output: Clear Decisions and Reports

right and this is an end -to -end workflow because now you can generate a report in say English for now and send it off to the person who asks the query so this is

Adoption, Reliability, and User Responsibility

how AI is changing how people work is it everywhere no of course not think these things like this take time for adoption right it's a whole different way of working it's a whole different way to even understand when the AI might be

giving you complete bullshit sorry for the word language but you know it is it is because AI sometimes can tell you the most confidently that it knows what it is doing but no it doesn't right most tools have guardrails like ours we really emphasize on being correct but it is also upon the user to understand that

to gain a sort of intuition to say okay this is reliable I move forward with this what took me three hours before takes me five minutes but if it is incorrect, it takes me another 10, 15 to correct it, right?

Real-World Use and Time Savings

So that is the challenge, but we see a lot of people actually working with this, using this on a day -to -day basis.

Conclusion

I could talk about it a lot more, but given that we're out of time, I'll stop for now.

Thank you so much.