The models were good in a notebook and useless in production, where nothing was reproducible and every handoff lost something. I built the path from experiment to serving that actually held up — versioned data, a real evaluation gate, and deploys that could be rolled back without ceremony.

Reworked features and validation; offline accuracy rose from 35% to 66% and held online.

Everyone wanted 'AI in the product' and nobody had said what good would look like, so we were one demo away from shipping something confidently wrong. I set the evaluation first — the metric, the holdout, the bar to clear — so we could tell real progress from a nice-looking demo.

Moved training to preemptible hardware with safe checkpointing; compute cost fell 55% at the same wall-clock.

Labeling was slow, inconsistent, and the bottleneck on every model we wanted to ship. I built the tooling and the guidelines that made it fast and repeatable, and put the annotator disagreements to work as a signal instead of noise.

Built an A/B analysis layer; 8 product teams now read results without a data scientist in the loop.