← Back to Resources

Monitoring Models in Production

Practical Resources - AI Engineering

Deploying a model isn't the finish line — it's the point where a new kind of work starts. Unlike most software, a machine learning model can silently get worse over time without a single line of its code changing, purely because the world it's making predictions about has moved on. Monitoring is how you find out before your users, or your grade, do.


What to monitor — three broad layers:

The tricky part: you often don't have ground truth immediately. In many real systems, you don't find out whether a prediction was actually correct until much later (or not at all) — you might know a loan default happened months after the prediction, or never directly know if a recommendation was "correct." This is exactly why monitoring input data distributions and prediction distributions matters: they're available immediately and can act as an early warning sign, even before you have the labels needed to measure accuracy directly.

Setting up useful alerts:

Why is this important? A model that isn't monitored can fail silently for a long time, and by the time someone notices (usually because a person complains, or a downstream number looks off), real damage may already be done. Monitoring turns "we hope it's still working" into "we know it's still working," which matters a great deal once real users or real decisions depend on the model's output.

Where to go deeper: the open-source Evidently AI library is built specifically for ML monitoring — data drift, prediction drift, and model quality reports — and its documentation is a good practical entry point even for a smaller student-scale project.