I build the infrastructure AI agents run on, and the data work that runs on top of it.

A proxy in front of my agent’s tool servers. It compresses what they return and forwards the original untouched whenever a perfect round trip cannot be proven.

Given a diff, it names the user journeys the change could have broken. Everything comparable answers in tests and needs coverage first; this answers on a repo that has never run its suite.

Stops an agent writing a call to a function the repo does not have, before the write lands. No model, no network, ~12 ms — and a clean check costs no context at all.

Replays real bug-fix commits from my own repos against cheap models, several times each, and reports cost per solved task. It publishes its own corrections.

Local retrieval over years of my own working notes. It beat keyword search at a thousand documents and lost to it at fifty-seven thousand — so I published that.

Spatial models over Michigan well records, and the contractor-intake business built around them. Depth is predictable from location. Yield is not.