Two built at work, four built to answer a question I could not put down. Every case study carries the same anatomy — the constraints, the alternatives I rejected, the numbers, and the parts that still are not done.
6 · one template
2 · AXA Group Operations
4 · built on my own
2 · both free-tier hosted
[[projects-stack]]
Same six blocks, same order, six times. If you have read one of these pages you already know where to find the part you care about.
Problem, approach, stack and outcome in a table you can read in fifteen seconds. Read only this and you still know whether the rest is worth your time.
The one thing that turned out to be true and was not obvious at the start. Usually the reason the first approach was thrown away.
How the pieces fit, drawn rather than described, with the data path stated end to end.
What the system had to do, and what it was not allowed to do. Latency budgets, confidentiality, cost ceilings, the people who had to sign off.
The ones nobody wrote down and I found by hitting them. This is the section most worth your time if you are trying to judge whether I have shipped anything real.
Selected work
Six systems: one built at work, five on my own time. Every number below is measured and sourced.
End-to-end fraud scoring platform where every feature reads only a transaction's past, and each shortcut that reads its future is priced.
Sources. Public IEEE-CIS data, never deployed. PR-AUC interval 0.5274 to 0.5614, 95 percent bootstrap, 1,000 resamples of the test month.
Estimate incremental impact and turn it into a spend policy. Decision science rather than prediction: the question is not who will convert, it is who converts because you spent the money.
NLP system that maps audit findings to compliance controls, replacing a manual mapping pass that ran for weeks each quarter at AXA Group Operations.
Sources. AXA Group Operations deployment, signed off internally. Manual audit effort down 96%.
Leakage-safe financial forecasting, built point-in-time so a backtest reports what the model could actually have known on the day rather than what the dataset knows now.
Clause-grounded RAG for insurance policy questions. It answers across eight plan variants, flags when the answer differs between them, and lists the clauses it retrieved rather than asking you to trust it.
Monitoring service for models already in production: data drift, performance decay, and an alert that fires before a stakeholder notices instead of after.