329901-1536x2732-phone-hd-the-imitation-game-wallpaper-image_edited.jpg

My Data Science Projects

Behind every project is a question that needed answering, a system that needed fixing, or a pattern waiting to be found. This is where raw data meets real-world problems—and where I get to build, break, and rebuild until insight emerges. Each project here is a piece of that journey.

The system answering "Is maternity covered in the first year?", showing the retrieved policy clauses, a grounded answer that maternity is excluded, and a verified citation to section 10.1.r.

ClauseCheck: Insurance Policy Q&A (RAG)

Health insurance policies hold the answers people pay for, buried in hundreds of pages of legal prose that read the same whether they are right or wrong. ClauseCheck answers a plain-English question with the exact clause it came from, refuses when the policy has no answer, and verifies every citation instead of trusting it.

FinSight: Leakage-safe financial forecasting

Leakage-safe forecasting of next-quarter Revenue, Operating Income, and Free Cash Flow for 117 US-listed companies — point-in-time SEC EDGAR + FRED data, five models compared, best served per target.

Minimalist line illustration on an off-white background. Two curves run almost on top of each other, then separate at a single point marked with a small neon yellow-green dot.

Drift-watch: ML Model Monitoring Service

A deployed model lives on two clocks. One tells you when its inputs change. The other tells you, weeks later, whether it was ever right. Drift-watch is built around both.

Dark hero banner reading "Fraud Detection" with the tagline "Ranking rare fraud under human-review capacity" and four stat tiles: 435 RPS on a single core, 28 ms p99 latency, under 1% fraud base rate, and gated retrain. A small chart shows a ranked fraud-score curve with the top slice shaded as the analyst review budget.

Fraud Detection Platform

Payment fraud hides in under 1% of transactions, and every alert costs a human analyst's time. This platform is a from-scratch recreation of a real-time fraud system I built at AXA. I rebuilt it around a sharper framing: fraud is not a classification problem but a ranking problem, bounded by how many alerts a team can review in a day. That one shift moves the decision threshold, not raw model accuracy, to the center of the product.

Automating Security Framework Mapping

A two-stage retrieval and ranking system that replaced a slow, inconsistent manual process with a defensible, explainable, production-adopted model.