Production Engineering on an ML Model

A track of P43 · Cross-Discipline Capstones.

Take a trained Data Science model (P5) and apply the full production discipline to it - package it (P35), gate it in CI on a quality floor (P36) and make its serving observable (P37) - the engineering around a model, not the modelling.

The discipline's flagship transfer task. Every other Advanced Python pillar builds a skill in isolation; this track proves the skills transfer — you take a real artefact from another discipline (a P5 tabular model) and engineer it to production standard end to end, without touching the modelling.

The model is given (a deterministic stub for a fitted pipeline), and both production simulators are inlined, so the whole capstone runs in Pyodide. You package it (a versioned manifest + a stable predict entry point — P35), CI-gate it (a pipeline whose quality gate fails a model that drops below a holdout-accuracy floor, so a regression can't ship — P36), and make its serving observable (structured prediction logs with inputs redacted + latency/volume metrics — P37). The capstone project then assembles all three into one service that's also resilient — idempotent by request id, so a retried prediction is served exactly once. It's graded by hidden tests and a design rubric, and it appears in the Projects gallery.

given model (P5)a fitted artefact, not rebuilt herepackage (P35)manifest + stable entry pointCI-gate (P36)fail below a holdout-accuracy floorobserve (P37)prediction logs + metrics
Production engineering applied to a Data Science model, end to end.