Review this Python inference handler that loads a model trained with a fitted scaler.
Offline metrics are strong but online predictions are nonsense. Find the train/serve skew bugs.