Pipelines

• Orchestrated training pipelines and idempotent, reproducible runs
• Feature stores and preventing training/serving skew
• Experiment tracking, model registry and artefact versioning

Serving

• Batch, online and streaming inference, and choosing between them
• Latency and throughput optimisation: batching, quantisation, distillation
• Autoscaling, cost control and the economics of GPU serving

Operations

• Data drift, concept drift and performance monitoring with actionable alerts
• Shadow deployment, canary release and safe rollback
• Incident response when a model degrades silently rather than failing loudly

Governance

• Model and data lineage for audit and reproducibility
• Access control, secrets handling and supply-chain risk in ML dependencies
• Retraining cadence, deprecation and retiring a model responsibly