The Workflow Is a Loop

Models are one stage in a system

A practical ML project moves through problem framing, data collection, exploration, splitting, preprocessing, baseline construction, training, evaluation, deployment, and monitoring. It is a loop because production evidence reveals new failure slices, labeling gaps, and changing conditions that feed the next iteration.

decision -> data -> split -> prepare -> baseline -> train
    ^                                      |
    └---- monitor <- deploy <- evaluate <---┘

Each stage should produce an artifact another person can inspect: a target definition, data contract, split rule, fitted pipeline, metric report, model version, and monitoring plan. A notebook with hidden state may support exploration, but a repeatable script or pipeline must recreate the final result from declared inputs.

Scenario: A model file performs well, but nobody saved the encoder used to create its 42 features. The weights are intact yet unusable because the input coordinate system is gone. The deployable unit was always preprocessing plus model plus schema.

Tip: Change one major assumption at a time and record it. Otherwise a better score cannot be attributed to data, features, algorithm, or chance.