MLOps

A model that produces good results in a notebook is not a product. Between that state and a system a business relies on sits the work MLOps describes: reproducible training, versioned models and data, a deployment that needs no manual steps, and the ability to withdraw a version when it turns out to be worse.

What separates this from classic DevOps is that here not only the code can change, but the world. A model can stay flawless and still get worse because the incoming data has shifted - different customers, different phrasing, a new product in the range. Without monitoring you notice only when somebody complains.

So measurement is part of shipping from the start: how often does the output deviate from what is expected? How does the input data change over time? And what does a run actually cost? Those three figures decide whether an AI system still holds up after a year.

For smaller applications MLOps does not necessarily mean a platform of its own. Often a clean pipeline, a handful of metrics, and the ability to return to the previous version with a configuration change are enough. What matters is that those paths exist before you need them.