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.

Instrumented live predictions for drift and degradation on 7 models; problems surfaced on a dashboard, not in support.

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.

Scheduled retraining with drift detection and a promotion gate; the model stopped silently going stale between releases.

We were flying blind on what the model did in production, with no idea it was wrong until a user complained loudly enough. I added monitoring on the predictions themselves, so drift and degradation showed up on a dashboard instead of in the support queue.

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

Retraining was a manual ritual only one person knew, so the model quietly went stale between the times they had a free afternoon for it. I automated the loop — scheduled retraining, drift detection, and a gate that only promoted a model that actually beat the one already live.

Added dbt tests and freshness SLAs; data-quality incidents fell 45% in a quarter.