geo3d
Rebuilds real places as verified, lightweight 3D scenes that run in a browser, from free public data. The reference build is a mountain valley; a second scene covers Gisenyi on Lake Kivu, in Rwanda.
Is it true in the real world?
The Lab
A lab that measures.
Beyond the products, the lab works on two things: grounding AI in the physical world, and making intelligence cheaper to run. Both follow one rule: validate a source before you trust it.
The Survey. A survey passes over the terrain. Ground it has measured is drawn in firm ink; the rest stays a faint guess.
How it works
01
Gather
Start from free public data.
02
Register
Align every layer to the most trusted measurement available.
03
Reject
Test each source against the ground and drop what fails.
04
Ship light
Encode it so it loads on ordinary hardware.
What it does
Research and reference builds
Checked against the ground
Every layer is registered against the best available measurement and its error is reported, not assumed.
Rejecting bad data
Candidate datasets are tested against ground truth before they are used. Ones that fail are left out, however convenient they are.
Light enough for a phone
Scenes stream in pieces as plain static files and run in a phone-class browser.
Smaller models
Compression and quantisation research aimed at capable language models that fit in very little space, for devices that are offline or on slow connections.
WarpLab
An open-source tuner for GPU code, in early beta: edit, compile, validate, benchmark, and keep a change only if the result is still correct.
What it is not
- Research and reference builds, not products you can buy today.
- Accuracy depends on the data a region has. Where public data is coarse, the result is coarser, and we say so.