I build data systems
that hold up to scrutiny.

A pipeline and lookup tool over the openFDA adverse-event database. It flags drug–reaction pairs worth a human looking at. Every score is a closed-form statistic over a 2×2 table, and each one carries its own bias diagnostics.

• BCPNN and MGPS disproportionality scoring with empirical-Bayes shrinkage. Five hyperparameters fitted by MLE across all 33,852 cells, so the prior comes from the data.
• A known-answer harness of 9 positive and 4 negative controls. It reproduces the statin–rhabdomyolysis reporting odds ratio of 12.95, and it picked out cerivastatin, withdrawn in 2001, as the highest-scoring statin without being told to.
• An active-comparator diagnostic that separates molecule effects from delivery artifacts. Exenatide device leakage drops from 36.23 to 1.01 against the right control. The signal was the injector, not the drug.
• Runs as a containerised Flask service with PostgreSQL, Alembic migrations, accounts, email verification, and a weekly digest on GitHub Actions.

Detecting epileptic seizures in scalp EEG for patients the model has never seen. The classifier was never the hard part. Building an evaluation you can trust is, when positives are three in a thousand and no two brains look alike. The live app runs the whole pipeline in the browser on a bundled recording, so there is nothing to upload.

• Leave-one-subject-out cross-validation across 23 patients, reaching 0.31 PR-AUC against a 0.004 baseline. Getting the protocol wrong is worth anywhere from −0.10 to +0.40 depending on the model, so there is no one number to quote for a dataset.
• Two metrics that quietly flatter a model. Averaging precision across folds read 0.23 where the pooled figure was 0.004, so metrics are pooled from summed confusion counts now. And a DummyClassifier that alarms almost constantly still scores a respectable 0.89 alarms/hour, because that metric counts alarm runs rather than time spent alarming.
• Decision threshold and run-length cannot be tuned separately. Chosen one after the other, the pair detected nothing at all. They are searched jointly now.
• Three upstream data bugs that corrupt labels without throwing an error: a library silently renaming duplicate channels, one mislabelled seizure end time, and an annotation field that turns out to be a byte offset rather than a timestamp.

A match prediction engine for top-tier football. Elo supplies the outcome probabilities and a Dixon-Coles grid is rescaled onto them, so the scorelines, expected goals and over/under numbers can never contradict the match result.

• Benchmarked Elo, Dixon-Coles and XGBoost on a held-out season, picking on log loss with paired bootstrap tests instead of point estimates. A 46-feature XGBoost turned out to be statistically indistinguishable from a 2-parameter Elo.
• Blending bookmaker odds adds nothing. Linear and logarithmic pooling both selected weight 0.00 on validation. The odds are still in the UI, they just do not touch the prediction.
• Serves from a FastAPI backend with a Vercel frontend, refreshed by a GitHub Actions job that rebuilds Gold data and commits it back to trigger a redeploy.

A copilot that scores how well a role fits a real CV, shows the gaps, and drafts the application. It does not scrape job boards and it does not submit anything on your behalf. The job market has a volume problem already. The live link opens a sign-in page because this is a real multi-user app with per-user isolation and cost caps, not a shared demo.

• One LLM interface over Anthropic, Gemini and Ollama, so the app moves between hosted and fully local models without touching a call site. Useful when a free tier quietly drops a capability on you.
• Job-fit scoring runs on local sentence embeddings, with a lexical fallback for when the model is not there. The hosted demo runs the fallback.
• The product constraints are five tests. The suite fails on a banned scraping library, a restricted-platform URL, a submit-shaped route, or a model call inside the quality gate.
• One test counts the outbound call sites in the discovery package and asserts there is exactly one. Another pins the allowlist that site checks against.
• Parses and rewrites résumés across PDF and DOCX with layout introspection, exporting ATS-safe documents.