M19 · Level III — AI Systems Engineering
ML Systems, MLOps & Inference
Engineer the operational life of intelligent models: packaging, serving, inference, observability, deployment, scaling, updates, rollback and production evidence.
Topics
Technical concepts
- 01The Model Artifact
- 02Training-Serving Parity
- 03Deterministic and Reproducible Inference
- 04Inference APIs as Typed Contracts
- 05Serving Architectures
- 06Latency as a Distribution
- 07Throughput and Saturation
- 08Dynamic Batching
- 09Inference Caching
- 10Memory as a Production Constraint
- 11Numerical Precision at Inference
- 12Model Registry and Provenance
- 13Reproducible Runtime Environments
- 14ML CI/CD Pipeline
- 15Offline/Online Parity
- 16Deployment Strategies
- 17Rollback Engineering
- 18Capacity and Autoscaling
- 19Backpressure and Load Shedding
- 20Observability: Logs, Metrics and Traces
- 21Service Level Objectives for ML Systems
- 22Drift and Change Detection
- 23Online Features and Data Dependencies
- 24Distributed Inference
- 25Queueing Reasoning for Inference Services
- 26Cost and Resource Efficiency
- 27Configuration, Secrets and Runtime State
- 28Controlled Change with Feature Flags
- 29Load Testing
- 30Soak Testing and Long-Run Stability
- 31Production Fault Injection
- 32ML Runtime Security
- 33Model Supply Chain and Integrity
- 34Inference Optimisation Before M20
Practical implementation
- 01Laboratory I: Observable Inference Service
- 02Laboratory II: Latency/Throughput Benchmark
- 03Laboratory III: Dynamic Batching Study
- 04Laboratory IV: Cache Correctness and Performance
- 05Laboratory V: Canary Deployment Simulator
- 06Laboratory VI: Shadow Evaluation
- 07Laboratory VII: Autoscaling and Backpressure
- 08Laboratory VIII: Rollback and Recovery Drill
- 09Laboratory IX: ML Production Incident Analysis
Tools and technologies
Project
End-to-end deployed service + evidence + rollback + monitoring
Expected learning outcomes
- 01Can package and reproduce a model artifact.
- 02Can define typed inference contracts.
- 03Can measure latency, throughput, resource use and saturation.
- 04Can reason about batching, caching and memory.
- 05Can deploy through controlled release strategies.
- 06Can instrument logs, metrics and traces.
- 07Can detect offline/online drift or parity failures.
- 08Can enforce backpressure and resource boundaries.
- 09Can execute and verify a rollback.
- 10Can investigate an ML production incident from evidence.
Assessment
- 01Daily practice
- 02Laboratories
- 03Research reproduction
- 04Integrated project
- 05Technical viva
- 06Engineering record
Source document: TUTOR Frontier Module 19 ML Systems MLOps Inference