Production Cases

Here is the brief: a model in a pickle file, a server to put it on, and a world that will not stop drifting underneath it. Keep it alive for a quarter. Concretely — you inherit the income model from P5's tabular case, a fitted scikit-learn Pipeline that predicts whether a census respondent earns over $50K, and it stops being a notebook exercise the moment someone downstream starts consuming its scores. Every stage you need has already been taught; what has never been asked of you is doing them all, in order, on the same artefact, without dropping it between stages.

That is what a case is. The P13 tracks teach packaging, serving and monitoring each in isolation; a case picks up one real model and carries it package, serve, monitor — ending in a graded Project on the full 32,561-row census table. The second case repeats the discipline on a different species entirely: the TinyBERT checkpoint from P10, packaged and batch-scored inside the kernel, with an input-drift decision on live text. Same lifecycle, different failure surface — tabular drift announces itself in PSI on feature columns; text drift hides in an embedding distribution.

The reason cases exist is that production breaks at the joins, not in the stages. The checksum you pin while packaging is the thing monitoring later trusts; the eval baseline you record before serving is the only thing that makes a drift score a decision rather than a number; the model card you wrote is what tells the person paged at 3 a.m. what "normal" looks like. Run the joins yourself, end to end, twice, and the pillar's material stops being four separate skills and becomes one job.