Science taught me evidence. Trading taught me risk. Running a business's data taught me what actually moves it. Now I build AI systems, and science tools, that prove what they claim.
| Skill | Where it shows |
|---|---|
| Agentic AI & multi-agent systems | Tri-AI: planner, builder and reviewer agents; stuck-agent detection; failed work reverts itself |
| LLM evaluation & guardrails | Called all 574 models a router listed; 20 answered as themselves. Work counts only when a test exits 0 |
| RAG & grounded answers | Genclarus: hybrid BM25 + local embeddings, Recall@10 0.950; refuses any sentence it cannot cite |
| AI-assisted development | Claude Code and Codex daily: one AI writes, a second reviews, the test suite decides, I check the claims |
| Machine learning | Leak-free splits, honest baselines, browser inference matching the trained model to 7.5e-08 |
| Python, APIs & testing | FastAPI, Cython, SQL; pytest, Vitest, GitHub Actions; ~2,500 tests gate Tri-AI deploys |
| Bioinformatics | GRCh38, ClinVar, gnomAD, UniProt, Biopython, Kallisto, Sleuth |
Every project publishes what was tried and did not ship next to what did (see the DECISIONS.md files). When an earlier version claimed more than it measured, the fix is documented too.







