How do you debug a Keras model that trains well but performs poorly in production?

This is usually caused by data distribution mismatch, preprocessing mismatch, or overfitting to training pipeline artifacts.

Training vs production mismatch occurs when real-world inputs differ from training data assumptions. Common causes include inconsistent normalization, missing feature engineering steps, label leakage during training, or drift in feature distributions. Debugging requires comparing training pipeline vs inference pipeline step-by-step.