Data engineer with a PhD in computational biology. I use AI agents, mostly Claude Code and Codex, on open science problems and publish the code and the results. Results are unrefereed unless a paper says otherwise.
One hub per field. Each collects the work, the code and the references it builds on.
- Biology: ai-biology, protein design, the Arc Virtual Cell and Virtual Embryo challenges, open drug-target discovery, and refereed bug reports on scientists' published software.
- virtual-cell-challenge: pipeline and local evaluation harness for the Arc Virtual Cell Challenge 2026. Final entry submitted, scores pending.
- Virtual Embryo Challenge: competing now. The reusable parts are public in virtual-embryo-community: a stage-window checker for atlas sample sheets, and a best-of-k selection tutorial.
- adaptyv-challenges: entries to the Adaptyv/Anthropic protein design competition, challenges 3 to 5.
- louis: pick an autoimmune disease and get trust-ranked CD4+ T-cell drug targets from a genome-scale Perturb-seq screen.
- autoregulation: gene regulatory network stability analysis, code from my PhD (Feldman lab, Stanford).
- Pull requests to AI-for-biology companies: bug fixes to the public code of Edison, Latch, Chai, Rowan, Seqera, Arcadia, Adaptyv, Biomni, Robocurve and Synthesize Bio. Each bug was found by one model and confirmed by a referee from a different model family. 21 are open and none merged yet; the live table is in ai-biology.
- Math: math-certificates, an index of explicit, exactly checked bounds on open optimization constants. Each entry states its verification level: AI review only, an independent checker, or Lean.
- Six of these are merged into Terence Tao's optimization-constants table: C86 ≤ 1.8117251, C82 < 0.2534449, and C68, C42, C44 and C70. Seventeen more pull requests are open, several of them improving the merged ones.
- Public postings on Hexagon: C86, C82 and Korenblum's c2, plus later bounds for C44, C86 and C68.
- Physics: ai-physics, certified bounds on published results, exact checks with independent referees, and bug reports on physics software.
- Astronomy: planet-hunt, AI agents searching NASA TESS data for small planets. Four candidate-planet research notes went to RNAAS. Three were declined under the journal's new policy on notes about transits in a few stars, and a revised note on the TOI-4342 candidate is under review.
- family-learning-games: 13 self-hosted learning games and an illustrated story book that writes a new chapter every night, in one hub.
- rstack: the stack I run research agents on. A research harness with checks against overfitting, stalled work and unapproved outreach; a file-based knowledge base the agents maintain; and the guards, scheduled jobs and lints that keep it honest, with tests. Point Claude Code or Codex at it to install.
- sentinel: two-stage incident detector for open-source repos, built with Codex.
- prediction-market-mm: a passive market maker for the Orderbook Prediction Market Challenge, plus the harness used to tune it.