I build backend systems and data pipelines, then make them decide things — with retrieval and LLM judgment where that's actually the cheapest path to a correct answer.

A deterministic ETL and scoring engine that turns France's public legal-announcement feed into prioritized, auditable leads.

Batching database commits instead of committing per record turned a 977.5s production run into 58s on an identical re-run — a 17x wall-clock improvement, found by benchmarking rather than guessing.

A focused web crawler that decides what's worth fetching before it fetches it, cutting a 600-second crawl to 15.

Same 560-node benchmark workload: ~600s to ~15s wall-clock (~40x), by cutting LLM calls to about 1% of link volume instead of exhaustive traversal.

A CLI that turns semi-structured invoice PDFs into validated, typed records — no OCR, no LLM, no database, deliberately.

200+ regression tests across table and paragraph invoice layouts and regional number formats, with zero external service dependencies.

Information retrieval built from first principles — analyzers, inverted indexes, and Boolean querying, not a wrapper around a search library.

A clean separation of ingestion, analysis, indexing, querying, and evaluation — each concern swappable without touching the others.