The gap between a model that works in a notebook and one that runs reliably in production is mostly engineering, and most of that engineering is unremarkable — versioning, reproducibility, deployment, monitoring. The industry's tooling is built for scale that few teams have.
Start with reproducibility, because everything else depends on it. That means the training data is versioned, the code is in version control, the dependencies are pinned, and the random seeds are fixed. If you cannot reproduce last month's model, you cannot investigate why this month's is worse.
Track experiments somewhere other than filenames. A lightweight tool such as MLflow, or even a disciplined spreadsheet at the very beginning, records which parameters produced which metrics. The failure mode without it is a directory of files named model_final_v3_actually_final and no memory of what distinguished them.
Treat model deployment as ordinary software deployment. The model is an artefact with a version, served behind an interface, released through the same pipeline as everything else, and rolled back the same way. Teams that invent a separate deployment path for models end up maintaining two systems and trusting neither.
Monitoring for models needs one thing beyond normal application monitoring: input distribution. Models degrade quietly when the data arriving stops resembling the data they were trained on, and no error is logged. Comparing the statistical profile of recent inputs against the training set catches this before the business notices.
Automate retraining last, not first. Scheduled retraining on a pipeline nobody is watching can silently ship a worse model. Manual retraining with a human comparing metrics before promotion is entirely respectable, and it is the right stage for most teams for longer than the conference talks suggest.
