Wrapping a model behind an endpoint
Model serving means taking a trained model and putting it behind an interface that other systems can call to get predictions, most commonly an API endpoint. A request comes in with input data, the model runs inference, and a prediction goes back out. This sounds simple, but it means the model now has to behave like a piece of infrastructure: it needs to start up reliably, handle concurrent requests, validate its inputs, and fail gracefully instead of crashing the whole service when something unexpected arrives.
The serving layer is also where a lot of practical decisions live that have nothing to do with model accuracy: how the input gets preprocessed to match training-time feature transformations exactly, how the output gets formatted for the calling application, and what happens if the model is slow to respond or unavailable.