Independent researcher working on evaluation and measurement for ML systems — figuring out whether a model or method actually does what it claims, and saying so honestly when it doesn't.
Everything here follows the same rule: the plan is frozen before the run, negative results are reported next to the positive ones, and every number reproduces from a single command.
A few things I've built:
- 🧪 llm-jury — six LLM judges vote on whether a RAG answer is hallucinated; together they beat any single one, and they caught real labeling errors in the benchmark itself.
- 📈 driftbet — pinpoints what caused a data stream to drift (model, noise, the world, or bad labels) without labels, with statistical guarantees that hold at any stopping time.
- 🔍 whest-teardown — reproduced a published estimator; a plain method beat it ~9×, and I found an accounting hole in its scoring.
- 🔭 Plus a pre-registered null result on real Kepler photometry and a cross-model study of LLM belief revision — the same rigor, different fields.