Distributed training platforms, real-time inference under strict SLOs, drift detection, and evaluation systems. The discipline comes from a background in physics instrumentation R&D at NASA JPL, Fermilab, and Berkeley Lab; the scale from senior ML engineering at Nubank; the range from founding and shipping a production RAG platform solo.
Four times I measured a real improvement to a spacecraft anomaly detector. And the model was never one of them.
Read MoreExtending the ESA telemetry platform to a live ISS relay feed — and discovering that the assumptions baked into a clean, archived benchmark quietly break the moment the data stops being continuous.
A self-correcting text-to-SQL agent that detects its own performance drift with windowed statistics and learns from its failures — recovery validated by a McNemar test. I owned the drift-detection stage.
An end-to-end MLOps platform detecting anomalies across hundreds of channels of real ESA spacecraft telemetry — built so per-channel models stay trainable and servable without infra cost exploding.
A multi-tenant reviews-to-answers system — a RAG service, a Next.js ordering platform, and a Solana review program tied together on-chain. My strongest example of production infrastructure thinking.