Launch HN: Bild AI (YC W25) – Understand Construction Blueprints Using AI
The problem we're tackling is the sheer manual effort that goes into generating material quantities and cost estimates from blueprints today. Contractors and suppliers spend countless hours doing takeoffs by drawing on blueprints by hand - it's tedious, error-prone, and costs the global industry $30B a year. A single mistake can lead to thousands in losses on a project.
My co-founder Puneet experienced this firsthand as he was building hundreds of houses in Canada. Meanwhile, my background is in applied ML - I started at Google at 19, then Waymo where I built perception models for self-driving cars. Puneet and I met at Hack for Social Impact where we built our first _very_ narrow-scoped prototype.
Since then we’ve expanded our scope slowly, with a laser-focus on accuracy. Our approach is to use a suite of specialized machine learning models for specific blueprint comprehension tasks, rather than a single end-to-end model. For example, we've developed computer vision models that are highly accurate at detecting and measuring floor areas, or identifying and counting framing elements like studs and doors. By composing these expert sub-models, we can achieve high accuracy on the overall takeoff.
This is somewhat analogous to the approach we took at Waymo for self-driving perception - having an array of dedicated models for tasks like lane detection, traffic light classification etc. It's a very different paradigm than the big-data end-to-end models like what Tesla uses - unfortunately we just don’t have enough data yet.
We're working with some early customers like flooring suppliers to help automate their estimating workflows. But we see a huge opportunity to expand this "AI that understands blueprints" approach across all trades.
Would love to get the HN community's thoughts and feedback! Construction is an industry I think is really ripe for applying cutting-edge ML techniques. If you have experience in this domain as a builder, architect, estimator, supplier etc. I'd love to hear about your workflow and pain points.
Also if you're a researcher or engineer excited about applying state-of-the-art techniques to real-world problems in underserved industries, definitely reach out!
We’re currently live with customers but are only able to serve a subset of trades accurately right now. Head to https://www.bild.ai/upload if you want to try uploading a blueprint and we can talk about your use-case. I truly believe we can solve blueprint understanding with AI, and there seem to be a huge number of applications. I'll be here all day to chat and answer questions!
41 comments
[ 4.4 ms ] story [ 81.1 ms ] threadIt looks like your launch is opening this up to the general public - why not niche down to GCs? Maybe the launch is focused on simply gathering more blueprint data to feed your models?
The benefit of estimating quantities and cost cycles in with pre-con and business development, the artifacts during the pre-con design phase tend to be different than the takeoff artifacts which are often transformed through BIM.
Did you learn something to the contrary? Or are you purposely targeting smaller firms and projects that don't use Bim and maybe won't for a long time?
[1] https://en.wikipedia.org/wiki/ISO_10303
BIM and other standardization is really the correct answer to this problem. This is a stop gap to cover for when/if that ever gets widely adopted.
On the other side, architects are using Revit more and more and takeoffs like square footage of flooring are accurate and take no time at all. That's another industry slow to change and that used to take more effort so many architects aren't providing that information to their clients, but technically there's nothing preventing it. There's a bit more hand waving when it comes to calculating number of studs etc, but that is pretty straightforward as well.
Source: I'm funemployed as a drafter for a local architect after 25 years in software.
As a side point - sometimes the shady lumberyards do bid too low on purpose to win business. Then later have the contractor submit a change order. This often hurts their reputation unless the contractor is in on it to win a bid. The supplier doesn’t tend to lose money though as the bid is for the quantity of materials.
You could turn this into a selling point. As in, helping a contractor or competing supplier verify the takeoff.
I agree with your point, having a second, impartial source is important to confirm the ballpark.
Here are things to consider:
Experienced builders don't care about the takeoffs on a big picture basis, the takeoffs are usually wrong, even if perfectly done. In our experience half of drawings we receive, are heavily revised by the order is approved (heavily revised meaning over 10% has changed). EWP, structural metal need to be accurate but framing lumber, and sheet good can be off on counts at the lift quantity (+-1 lift for an average house).
Suppliers aren't responsible for the takeoff so the amount the quote is negligible (see drawing revisions, and trades can misallocate the materials - This can't be reasonably traced). Over? The customer ends up paying less, under? The customer pays more. This has been universal where I am (Ontario, Canada).
A large minority of plans are missing key elements (like sheer walls), pointing out, and showing these differences would be a big value add for the consumer (contractors using the materials) by the supplier.
Good customers understand that lumber is a commodity, a lower price this week can flip next week, and they'll contact their preferred vendor about the differences.
There's always a preferred vendor.
Not great customers will shoot drawing off to multiple suppliers, causing them all to do the same takeoff, wasting time, and money, only to deal with the same issues above. They'll still go back to their preferred vendor to get the lowest price.
Summary of the above is: EWP, and structural metal are key items because they rarely change, framing lumber, and sheathing requirements change all the time. What you're looking at is helping suppliers capture the bad customers (which are often the biggest, to be clear), but saving suppliers the time handling them is great. Also, accuracy, and pricing isn't that important (with caveats).
This isn't a statistically significant sample size, consider it anecdotal.
Estimators miss things ALL THE TIME. It's the subject of seemingly endless in-house arguments between PMs and Estimators:)
I'm asking because even though I am (mostly) technically illiterate I have asked both ChatGPT and Claude to help me build a scraper for construction material costs, from the suppliers we use, that can be updated in realtime or at least monthly. Haven't done anything with those instructions yet, but I would love nothing more than to use a tool that we could feed a blueprint into and then would tell me, with "laser-focus accuracy" <smile> how many x's the project would need and the costs. Even better yet if it could compare costs from suppliers and guide us to the lowest-cost supplier.
Edit: oh, while you're thinking of replying, how high fidelity do the blueprints need to be? Again, I'm sure you specify somewhere, but too lazy to find it. How far along the spectrum from "drawn on a napkin" to "fully standardized" do you accept?
For the second question, it really is most accurate on "fully standardized" blueprints due to our training distribution. Will work on improving that as well!
Best of luck with the business (and with getting to know the corp dev people at Autodesk/Procore/etc.--sorry, couldn't help myself!).
You'd still need people to check what actually got installed, so that you can bill for it. Like, there's only so much you can determine off the plans.
And what happens if (when) the plans are wrong or impractical?
My Dad worked in construction for his career, and I did briefly, and there's generally a lot of stuff that needs to be figured out on site due to physical or logistical constraints.
Sounds like this is just for homebuilding though, which is a much easier problem.
Separately, it seems like it would be incredibly useful to use your models in various embodied carbon estimation tooling and other decarbonization research streams. Have you thought about partnering with any academic researchers on this? If you are interested, let me know, as I can definitely connect you with a bunch of researchers who would be interested!
What's crazy about this is that the AI revolution is going nuts. We've started with Steel and customers who would traditionally bid on paper are now jumping straight to AI takeoffs. The impact is real.
One customer recently told us that he was able to bid on $200M more than he would have been able to otherwise: https://www.linkedin.com/feed/update/urn:li:activity:7300899.... That's a couple of million in revenue that they would have worked away from because of capacity constraints.
One thing I've been thinking about is if you could use a model like this as the first pass for permitters (Like a GitHub Actions CI/CD) who review blueprints.
Many developers use the regulatory side of various engineering approval processes as a quality control check which costs money and time for the regulator who is tasked with enforcing a standard.
It would also be good to speed up the workflow for developers saying hey, this thing looks weird did you really mean to do this?
And then further on, you could add a way to check it for constructability. My framer friends often get annoyed at whatever engineer because the way the structure is designed is materially inefficient or hard to construct.
- You’re right, data is very hard to come by. I’m curious, how do you plan to get around this? Outsourcing human labeling? We found it to be a very difficult task.
- The subcontractors and local construction companies we talked to were overwhelming excited about the idea.
- It’s entire people’s jobs to get this done and done correctly. They sit on site holding the pdfs in their hands, manually counting and calculating. You bet a lot of mistakes occur. They would absolutely love to have a digital assistant for this.
- Some of them (especially managers and owners) are quite technical and are using software such as BlueBeam and other CAD software to make these calculations. It’s quite manual currently, but gives great insight into a better solution. This led us to having the user manually select the symbol they wanted counted (which ML struggled to get right). Just getting the part counts (and highlighting them in the pdf) was a huge help!
- Impressive you got square footage calculations correct! In our experience, there was way too much variation between architects (and multistep dimension labeling) which made it hard (even for humans) to get right. How has your model generalized OOD thus far?
- Are you planning to integrate voice? Many of the subcontractors we worked with are very low tech. They usually talk with their clients in person, on the phone, or maybe text. But they don’t use email or their smart phones for much.
I will be following your work! I have friends who would love to use this once it passes the human threshold.
Coincidentally, yesterday I had a client meeting and they ask for exactly that. I'm working as lead developer for https://howie.systems and we are building a co-pilot (knowledge platform) for the AEC industry.
Would love to have a talk. Your product could save us lot's of work!
If owners/developers understood this they could create contract structures that incentivize more fluid data collaboration aka the quantity take offs automatically generate as you are designing.
Pragmatically though in the current AEC landscape there is still a need for 2D QTO, nice work