17.2 Model Monitoring

A model that works on launch day can quietly get worse as the world changes — customers behave differently, prices shift, new fraud patterns appear. Model monitoring continuously tracks predictions, accuracy, latency, and input distributions so you are alerted before users feel the damage.

Consider a crop-disease classifier used by farmers around Bahir Dar. It was trained mostly on dry-season photos. When the rainy season arrives, leaves look different and accuracy drops. Monitoring detects this 'data drift' and signals that retraining is needed.