How do you build a custom sklearn transformer, and why...

You get there by inheriting from BaseEstimator and TransformerMixin and implementing two methods: fit(X, y=None), which learns whatever statistics you need from the data it’s given, and transform(X), which applies those already-learned statistics to any data handed to it. That separation is the entire point — it’s the same fit/transform contract that StandardScaler follows, so a custom transformer gets the exact same leakage protection automatically.

The reason this beats a standalone preprocessing function is composability: once your class follows the contract, it slots directly into an sklearn Pipeline alongside StandardScaler and your model. That means it gets cross-validated correctly (each fold fits fresh), gets saved as part of one joblib artifact for deployment, and can never accidentally be fit on the test split by a careless call somewhere in the code — the Pipeline structure enforces the discipline instead of relying on the engineer to remember it.