Take a working model from a notebook into a genuine, monitored production system serving real users
You can package and deploy a model behind a genuine API, choose appropriate serving infrastructure for different latency/scale needs, monitor a deployed model for both technical and data-drift problems, and design a safe rollback and retraining process.
Demonstrate genuine model-packaging/serving judgement, infrastructure trade-off awareness, drift-monitoring understanding, and safe deployment/retraining discipline.