MLOps maturity: a staged model for production machine learning

Models are built and run manually by individuals. Reproducibility depends on memory. This is appropriate for exploration but risky for anything relied upon.

Code is versioned, environments are specified and training can be rerun. Deployment is still manual.

Pipelines train, test and deploy models with minimal manual steps. Automated validation gates protect production.

Production data and predictions are monitored for quality and drift, with alerts and runbooks.

Lineage, approvals, documentation and risk classification are built into the platform. Organisations should target the stage appropriate to the risk of their models, not necessarily the highest.