The lab — 14 experiments, all live
Proof is something
you can touch.
Real simulations, solvers and models, computing on your device as you watch — and where we use sample data, a script or assets generated offline, we say so. Each runs on engines we have spent about ten years and millions of dollars building, written with our agents and reviewed line by line by senior engineers: the same loop we run for clients.
01 Fluid dynamics
The solver, opened up.
The Navier–Stokes solver that runs under the homepage, with the code taken away and its working shown the way an engineer reads it: dye density, streamlines at equal steps of flux, vorticity and pressure as contours, vectors for the speed. Switch any layer. The body in the stream is the constraint the solver enforces — the water cannot be compressed — and the readout measures how well it holds. Move through it.
- Running
- WebGL2 stable fluids · half-float textures · up to 24 pressure iterations a step · stream function summed on the GPU · divergence read back live
- For clients
- Flow, spread and congestion — water, air, crowds, capital — modelled, then opened up so the people who sign off can see why.
02 Chemistry
Morphogenesis.
Two virtual chemicals react and diffuse — the Gray–Scott model, heir to Alan Turing's 1952 theory of biological pattern. Coral, dividing cells, fingerprints and worms each come from just two numbers. Seed it, carve it, and watch it heal.
- Running
- GPU reaction–diffusion · up to 512² cells · 8–14 solver steps a frame · a live map of the parameter plane
- For clients
- Pattern formation in materials, biology and markets — and the intuition to know when simple rules create complex outcomes.
03 Machine learning
A model, learning.
A neural network trains from random weights to separate three interleaved spirals — a problem no straight line can solve. Then it shows how: carried through its layers, the curved boundary straightens into three lines. Add a data point and it adapts.
- Running
- Two hidden layers of 16 · tanh · Adam optimiser · backpropagation in plain JavaScript
- For clients
- We train, evaluate and monitor models on your data — and we show you how they decide.
04 Language models
Inside attention.
How a language model decides what “it” refers to. Swap one verb and, in layer 9, the pronoun moves from the transfer to the trustee. Each head has its own habit — position, grammar, meaning. Point at a word to see what it reads; move the temperature and all 144 heads follow.
- Running
- Illustrative logits (sample data) · live softmax across 12 layers × 12 heads · Canvas renderer
- For clients
- Agents that read contracts, mandates and case files — with the evidence for every answer.
05 Agent-based simulation
Emergence.
Thousands of starlings, each watching only its seven nearest neighbours. No leader, no plan — yet a turn sweeps through the flock like a wave. You are the falcon. This is how we simulate crowds, fleets and markets before anyone builds anything.
- Running
- Topological flocking, seven neighbours · spatial hashing in a Web Worker · instanced 3D rendering
- For clients
- Test the policy, the layout or the rota on ten thousand simulated days first.
06 Optimisation
Routes that untangle.
Five vans, one depot and a day of deliveries on real streets, across a river with three bridges. The solver drafts a plan, then untangles it move by move while you watch. Drop in a rush order and it re-plans in milliseconds.
- Running
- Street-graph shortest paths · sweep clustering · nearest neighbour · 2-opt and exchange moves · ruin-and-recreate
- For clients
- Dispatch, rostering, berth and shift planning — solved, and explained to the people who sign off.
07 Document intelligence
Paper in, ledger out.
A photographed invoice is straightened, read and checked: fourteen fields located, the totals recomputed, the entry drafted — and then a person signs it off. The supplier is fictional; the workflow is real.
- Running
- Perspective correction · word detection in a Web Worker · 14-field extraction · arithmetic checks · human sign-off
- For clients
- Invoices, KYC packs, claims and contracts processed in seconds, with an audit trail a regulator can read.
08 Generative imagery
From noise, an image.
Diffusion models start from pure noise and remove it step by step. Here the sampler chooses between six images from one random seed — watch it hesitate, then commit. Drag the scrubber to walk the schedule; roll a new seed for a new outcome.
- Running
- DDIM sampler · cosine schedule, 50 steps · exact denoiser over six images — no trained network · WebGL2
- For clients
- Product imagery, synthetic training data and visual QA — generated responsibly, labelled honestly.
09 Simulation at scale
Two galaxies, colliding.
Up to sixty-five thousand stars, each pulled by two galactic cores that swing past each other and merge — integrated on your GPU every frame. Tidal tails and bridges emerge from gravity alone, as Toomre & Toomre first showed in 1972. Drag to orbit.
- Running
- WebGL2 GPGPU integration · float textures · restricted three-body model · up to 65,536 stars
- For clients
- When the problem is big, we put it on the GPU — risk, logistics, physics, pricing.
10 Physics
Woven from code.
A sheet woven from its own source code — 2,560 characters hung from eight pegs, held together by constraints. Grab it, pull it, tear it. A wave of mending knits it back.
- Running
- Position-based Verlet · 64 × 40 nodes · up to 12 relaxation passes · tearing and mending
- For clients
- Interfaces and simulations that feel physical — training tools, digital twins, product configurators.
11 Voice agents
A call, handled.
A caller wants a parcel redirected. The agent finds the order, checks the new address and quotes the fee — and when the caller cuts in, frustrated, it stops talking and hands over to a person, already briefed. A scripted demo, built on the patterns we deploy.
- Running
- Streaming transcript · intent & entity detection · tool calls · sentiment hand-off rules
- For clients
- Contact centres that answer at 3am in any language — and escalate exactly when they should.
12 Geospatial
Terrain, eroded.
A landscape generated from noise, carved by forty thousand simulated raindrops, then flown over and drawn in glyphs and contours. Inspect any point, or make it rain.
- Running
- 256² fBm heightfield · 40,000-droplet hydraulic erosion · GPU raymarch · glyph renderer
- For clients
- Sites, catchments, corridors and assets — mapped, simulated and monitored.
13 3D generation
Described, then built.
Each of these five models began as a single sentence — its prompt is printed beneath it — and was generated, offline, into a textured mesh. What runs here, live, is the reading: a scan plane shows each one four ways — solid, wireframe, points and glyphs.
- Running
- Text-to-3D meshes, generated offline · glTF · live WebGL turntable · scan-plane transitions
- For clients
- Product visualisation, digital twins and training content at a fraction of the usual cost and time.
14 Computer vision
The lens.
A rain-wet street at dusk, from one of our project galleries, read the way a vision model starts to read the world: edges, density, direction — and beneath them the program that does the reading. Move the lens. Or point it at yourself: with your say-so, the same lens reads your camera live. The feed never leaves this tab.
- Running
- Sobel gradients · contour tracing · luminance-to-glyph shader · optional live camera, read on your own GPU
- For clients
- Inspection, counting and tracking from cameras you already own.
Every experiment, one loop
Built by agents.
Signed by engineers.
That is also how we will build your system — on engines we have spent a decade and millions of dollars building, with senior engineers on every line. Tell us what it is.