Chapter Seven · From Idea to Product
The ML Workflow End to End
You now hold all the pieces — data, training, model types, honest measurement. This chapter assembles them into the thing they actually are in a company: one workflow, from a vague executive wish to a running service that customers feel, and onward into the years of watching and retraining that nobody puts in the launch announcement. One project — predicting which Plateful customers are about to drift away — rides the whole road with us.
It starts, as these things do, in a meeting. Plateful's exec team has seen the churn numbers and delivers a sentence Iris will hear many times in her career: "let's use AI to reduce churn." Six chapters ago that sentence would have sounded like a plan. Now she can hear what it actually is — a wish — and she is about to learn the craft of turning wishes into projects that survive contact with reality.
Five topics walk the road in order. First, framing: turning "reduce churn" into a precise, buildable prediction task — the step where most ML projects quietly die. Then feature engineering: manufacturing telling columns out of humble raw data. Then the training tournament: what "we're experimenting" really means, and how it stays honest. Then deployment: the moment a file of numbers becomes a service the app can ask. And finally the truth that makes ML an operation rather than a project: models age, and someone has to watch them.