The ML workflow, end to end

Every project starts with a decision someone wants to make. 'Which customers are likely to cancel next month?' is a useful question. 'Use AI on our data' is not. Write down what the model will predict, who will use the prediction, and what they will do differently because of it.

Then ask: could a simple rule do this? If 'customers who have not logged in for 30 days' catches most cancellations, start there. A model has to beat the simple rule to be worth building.

Next you collect examples where you already know the answer: past customers and whether they cancelled. This is labelled data. You clean it: fix missing values, remove duplicates, check that dates and units make sense.

Then you build features, the columns the model learns from. 'Number of logins in the last 30 days' or 'days since last purchase' often matter more than which model you pick. Good features come from understanding the business, not from clever code.

You split the data, train a model on one part and measure it on the part it has not seen. The first result is rarely good. You go back, add a feature, fix a data problem, try another model. This loop is the real work.

Keep notes on each attempt: the data version, the features, the settings and the score. Tools like MLflow or a plain spreadsheet both work. Without notes you will not remember which run was the good one.

A model is only useful once it reaches the people making the decision. That might be a nightly list in a spreadsheet or a live API. After launch you keep measuring, because the data will change. We come back to this in the shipping step.