FastAPI's type annotations, Pydantic models, and lifespan context manager solve three common ML serving problems: loading expensive artifacts before requests start, validating inputs without manual checks, and returning useful errors when dependencies aren't ready. This tutorial walks through the patterns used in the flight disruption prediction API.
ROC-AUC tells you how well a model ranks flights by delay risk. It doesn't tell you whether a 40% prediction means 40% of flights actually get delayed. Those are different properties, and the second one matters when you're building a risk gauge. This tutorial covers isotonic probability calibration, SHAP explanations per prediction, ONNX export with a parity test, and why all three belong in the same ML lifecycle.
A rolling delay rate that includes same-day flights is leakage. A temporal split that shuffles randomly is leakage. In a flight delay prediction pipeline, leakage produces models that look great on held-out data and fail immediately in production. This tutorial covers the shared contract module, Spark window functions bounded to strictly prior days, schema assertions that catch banned features before training, and a temporal split that respects the direction of time.
Position bias makes raw click-through rate a misleading metric for search quality. Team draft interleaving weaves two ranked lists together so each algorithm gets equal position exposure, letting you count statistically meaningful wins per query without any relevance labels.
