[ Back to Build home ]

Everything left is physical.

Build Labs* is Build's research group. We study how AI agents work with the physical world: maps, parcels, terrain, regulation and drawings, the backbone of the built world. We publish the benchmarks, environments and findings that come out of that work.

A building takes years, hundreds of specialists and thousands of documents to get from idea to opening. Today's models were trained on text and code, but building runs on expertise that was never written down. Labs builds what's missing: the data, evaluations, modalities and models to transfer that work from humans to agents.

Most of what decides where and how things get built is not text. It is a flood map, a parcel boundary, a substation two kilometres away. Models are still weak here: they pick the wrong source, misread the map and lose track across the dozens of steps a real site assessment takes.

Build does this work every day for developers and investors, with experts reviewing every deliverable. That gives us real problems with checkable answers. Text, code, images and voice each had a team that took them seriously. We are doing that for the physical world.

* build /bɪld/

I. v.
To make, construct or create something physical or abstract by putting parts together over time.
II. n.
The structural quality or design of a manufactured item.
III. n.
A specific compiled release or update of a digital program.

* labs /læbz/

I. n.
A controlled environment where scientists study phenomena, test hypotheses, or manufacture chemicals and medicines.
II. n.
Any place or situation that acts as a testing ground for new ideas.

What we work on.

[ 01 ]

Data

Rebuilding the built world as data: finding, ingesting and structuring geospatial, planning and property data across markets, then turning real places into environments where agents can train and be tested.

[ Read more ]

[ 02 ]

Learning

Methods for learning from expert feedback at scale: capturing corrections efficiently, turning them into reward signal, and generating synthetic data grounded in real places.

[ Read more ]

[ 03 ]

Evaluation

Measuring what agents get right: verifying outputs against ground truth, calibrating judges against expert review, and publishing open-source benchmarks on complex, long-horizon tasks.

[ Read more ]

[ 04 ]

Modalities

Pushing what agents can do beyond text: reading maps and drawings, reasoning about space and geometry, and acting in the real world, from design through to construction.

[ Read more ]

[ 05 ]

Models

Building models that know the physical world: specialised by task, fine-tuned for each customer, and trained on the expert signals and evaluations that come out of real development work.

[ Read more ]

Projects and posts.

A blurred photo for a post still being written

[ Launching ]

Making the built world autonomous.

The hardest problem left in AI is the physical world. Want to help solve it, and transform the world around you? We're looking for:

  • Run your model on our benchmarks, or work with us on spatial environments for training.

    Email labs@build.inc