Projects

Auditable research provenance, fully local. Every number the tool reports traces back to the file that produced it — and it refuses to state more than the evidence supports.

Research generates artifacts: config files, eval dumps, benchmark tables. The numbers that end up in a README get transcribed by hand, and nothing checks them again. attestation closes that loop. It reads artifacts that already exist — no instrumentation, no log_metric() calls, no change to how anything runs — and answers questions like which arm of that sweep actually won, and on what evidence? and is what my README says still true?

Built as an MCP server so an agent can call it as typed tools rather than interpreting prose. Runs against local Ollama by default, or any OpenAI-compatible backend. One SQLite file, no new services.

A tool that reports success for work it did not do is worse than no tool.

A 150K-line computational chemistry engine, validated to micro-Hartree agreement against PySCF across 1,705 tests.

A return to the physics I trained on, rebuilt with modern tooling. Correctness in this domain is unusually well-defined — reference implementations exist and agreement is measured in fractions of a Hartree — which makes it an honest test of whether a system does what it claims.

Right-sizing Whisper to one deployment by calibrating on its own audio — no retraining.

Cutting four of whisper-small's 24 transformer layers removes 14.7% of its parameters and leaves word error rate essentially unchanged, for a single-speaker deployment. The fifth layer is a cliff, so the efficient frontier is sharp rather than gradual — and the two compression techniques don't compose: layer cuts and width pruning are each nearly free alone and break the model together.

Read the writeup →