ML Engineering Design

Designing the Systems Around Machine-Learned Models

A self-study course for working professionals who can train models in a notebook but want to design production ML systems end-to-end.

The model is not the product — the system is. This course teaches you to treat the system around the model as the primary engineering artifact: data & feature pipelines, experimentation, evaluation, serving, monitoring, and governance — with every decision made deliberately and recorded as a tradeoff.

What this course is

• Design-first, not algorithm-first. The emphasis is on architecture, requirements, and trade-offs — not on re-deriving ML math.
• End-to-end. It covers the full production ML lifecycle, including MLOps (CI/CD), deployment, observability, drift, and governance.
• Failure-driven. Every module exists to prevent a specific production failure mode (leakage, skew, drift, feedback loops, silent degradation).
• Vendor-neutral Python. scikit-learn + PyTorch, Pandera/Great Expectations, MLflow, FastAPI + Docker, Prometheus/Grafana. No single-cloud assumption.

What this course is not (yet)

• No labs — full coded demos are deferred to a second pass (💻 CODED DEMO (Pass 2) placeholders mark where they'll go).
• No assessment — deferred.
• No capstone decision — deferred (see M18).

Structure

18 modules in 6 parts, following the production ML lifecycle.

Part	Modules	Theme
0	M1 · M2 · M3	Foundations — the lifecycle, failure science, architecture & stack
1	M4 · M5 · M6	The Data & Feature Layer
2	M7 · M8 · M9	The Modeling Layer
3	M10 · M11	The Verification Layer
4	M12 · M13 · M14	The Serving Layer
5	M15 · M16 · M17	The Operational Layer
6	M18	Synthesis — a complete worked design

How to use this course

1. Read the modules in order — each builds on the last, and the failure-class → layer map recurs throughout.
2. Do the design exercise at the end of every module. They're paper-based; the point is the thinking, not the code.
3. Track your decisions in the tradeoff ledger discipline (and ADR format) introduced in M3.
4. When a module marks 💻 CODED DEMO (Pass 2), note the concept — the code comes in the second pass.

Reference stack

• scikit-learn
• PyTorch
• Pandera / Great Expectations
• MLflow
• FastAPI
• Docker
• ONNX / TorchScript
• Prometheus
• Grafana
• alibi-detect (drift detection)
• Prefect (orchestration)

Tech references

Specs, versions, requirements, and setup for the tools behind the stack — the reference shelf module lessons link out to. Working demos that exercise them land in each module's code/ folder during Pass 2.

• Airflow — promotion pipeline (M13) & retraining loop (M16), infra path
• Kubeflow Pipelines — same concepts, Kubernetes path
• TFX — same concepts, canonical component graph
• Prefect — orchestration concept, no-infra path
• alibi-detect — drift detection (M15, M16)

Status

All 18 modules are written. See syllabus.md for the full blueprint and conventions. The capstone decision (M18) is deferred, and full coded demos remain a second pass (💻 CODED DEMO (Pass 2)).