Work

Everything here has a run behind it.

Benchmarks, agents, inference and the products they turned into. Each card carries one measured result and a link to the code, the notes or the product.

Showing all 14

Agents · Product

Oct 2026

A game tuned by AI playtesters

A Handful, a VR physics game, went through 21 revisions scored each round by cold-play AI tester panels and scripted bots. The panel plateaued short of its target, and we say so.

panel 5.11 → 7.72target 8.5, not reached

Products · Agents

Oct 2026

Create

Describe a world, get a playable simulation: a language model writes a recipe from registered physics primitives, and a validator checks it before it runs.

20/20 playable in our eval~$0.29 per game

Archive

Earlier product work at NVIDIA, Lyft and Intuit

Five short case studies from the previous version of this site: enterprise AI catalogs, universal search, AV validation, and experimentation platforms.

Open the archive