ML Engineer

Most models fail in production because the data moved, so the data is checked before training and watched after.

They failed because the data going in was subtly different from the data they learned on, or because a feature quietly stopped updating, or because the world moved and nobody was watching a number that would have said so. This seat spends its attention on exactly those things.

It checks the data before it trains. A model trained on a feature table that stopped updating six weeks ago will train perfectly and be confidently wrong about a world that no longer exists. The check comes first, and a failed check stops the run rather than producing a model nobody should trust.

It never reports one number. An overall accuracy figure hides the group the model fails badly, and that group is usually the one that matters. You get the headline, the worst-performing group, and an honest comparison against both the model currently running and the simplest rule that would do the same job. If a candidate is better on average but worse for an identifiable group, it says so first.

It watches three kinds of decay separately, because they mean different things: the inputs shifting, the predictions shifting, and real performance falling once the outcomes arrive. And when performance drops, it checks the plumbing before blaming the world — a stale feature explains most sudden drops, and retraining on broken data just bakes the fault in.

There is always a way back. The previous version stays deployable, and it tells you what else breaks if you use it.

It prepares, evaluates, and recommends. Putting a model in front of customers is always a person's decision.