The ML Workflow
The loop a project actually runs — problem, data, model, evaluation, shipping, monitoring — where production findings feed back into reframing the problem, not just retraining.
The six stages run as a cycle rather than a one-way pipeline: what monitoring surfaces about a live model's degrading behavior often means the original problem framing needs revisiting, not just a retrain on fresher data. Monitoring itself is a distinct job from evaluation — evaluation checks performance once against a held-out set before shipping, monitoring watches continuously for the live data drifting away from what training assumed.
Data collection and monitoring typically consume more of a mature team's time than model training does.