01. Week 16 — Engineering Principles¶
🧠 Mental models¶
- One-way vs two-way decisions: "a concrete door versus a revolving door"
- ADRs: "the ship's log for why the course changed"
- Unit tests vs evals vs monitoring: "factory checks at build time, QA on the line, and alarms in the field"
- Technical debt: "interest charged on every shortcut"
- API-first: "rent first, build later if the neighborhood is truly worth it"
- Stakeholder communication: "the same engine shown through different dashboards"
⚠️ Common traps¶
- Arguing about tools before stating the real decision, constraints, and reversibility.
- Using unit tests to validate probabilistic model behavior that needs evals instead.
- Treating prompt or model-config edits as harmless text changes with no review trail.
- Choosing fine-tuning or self-hosting for prestige instead of evidence and operational need.
- Letting research-heavy work escape ownership, documentation, or rollback planning.
- Explaining architecture only in technical language and leaving product, finance, or legal unconvinced.
🔗 Prerequisites & connections¶
Builds on: Module 15 capstone trade-offs and the earlier modules where you learned modeling, evaluation, cost, latency, and system failure patterns.
Feeds into: Module 17 MLOps & Production, where reproducibility, versioning, release discipline, and incident response become operational requirements.
💬 Interview phrasing¶
- How do you decide between build, buy, prompt-plus-RAG, and fine-tune?
- What belongs in unit tests versus evals versus production monitoring for an ML system?
- Tell me about a one-way door decision you would treat differently from a reversible one.
- When is API-first the smart default instead of building in-house?
- How would you explain the same AI architecture choice to engineering, product, and finance?
⏱️ Difficulty markers¶
- 🟢 one-way vs two-way doors
- 🟢 ADRs
- 🟡 API-first vs build decision-making
- 🟡 unit tests vs evals vs monitoring
- 🟡 technical debt in AI systems
- 🔴 stakeholder tradeoff communication
Self-check questions¶
- What makes a decision one-way versus two-way?
- When is API-first the smart default?
- What exactly belongs in unit tests, evals, and monitoring?
- When should an AI workflow stay manual for now?
- What minimum documents make a system reviewable and operable?
- How would you explain a model choice differently to product and finance?
Health check¶
By end of Week 16, you should have: - [ ] Read 02_explainer.md completely. - [ ] Written at least one ADR. - [ ] Built one risk matrix for an AI workflow. - [ ] Distinguished manual, semi-automatic, and automatic stages for one feature. - [ ] Completed the hands_on_lab and revision. - [ ] Felt ready for 04_ml_platform_operations.