What you'll learn: a 90-day plan that follows this card: three projects of rising difficulty, a weekly design review, an incident-study habit, a short list of primary sources, and the evidence you keep.
At a glance
Figure 1. The weekly practice loop
How to read this: the loop repeats weekly; each pass leaves one written checklist change.
- Three projects in about 13 weeks, each harder than the last.
- One design review and one incident review a week.
- Expertise comes from repeated delivery; this plan starts the repetition.
Scenario
Composite scenario, not a real company. A typical reader, a backend engineer at a mid-sized firm, finishes the card and scores 11 of 14 on the blank-page defense. There is no AI project at work yet, and about five free hours a week (illustrative).
What you will be able to do
- Choose a project that targets your weakest criterion.
- Run a 60-minute design review on any design, yours or public.
- Turn a public incident review into a checklist line.
Prerequisites
- The blank-page design defense: its lowest criterion decides where you start.
- Advanced AI Architect lessons: LLMOps, reliability, security, governance.
Core ideas
The 90 days
| Weeks | Activity | Evidence you keep |
|---|---|---|
| 1 to 2 | Repeat your lowest rubric criterion on two new briefs (the optional labs) | Two scored sheets |
| 3 to 5 | Project 1 | Repository, golden set, evaluation report |
| 6 to 9 | Project 2 | Failure tests, approval demo, decision records |
| 10 to 13 | Project 3 | Threat model, trace sample, one-page briefing |
| Every week | Design review, incident study, one reading | Review notes |
Three projects of rising difficulty
| Project | What it proves | Done when |
|---|---|---|
| 1. A read-only assistant over documents you may use, with citations and refusal | You can ground answers and measure them | 30 or more golden questions, some unanswerable, with before and after scores |
| 2. A workflow that calls an API you control, with human approval and a log | You can make actions safe and test failures | Timeout, duplicate and bad-input tests pass; no write runs unapproved |
| 3. A multi-user system with access control, tracing, a budget and a small real pilot | You can operate what you design | A few real users for two weeks, a threat model, a tested rollback |
Keep projects small. Project 2 draws on reliability and scaling, project 3 on LLMOps and governance.
The weekly design review (60 minutes)
- Pick a design: a lab brief, your own project or a published architecture write-up.
- Run the seven steps for 30 minutes, with the threat-modeling lesson beside you.
- Defend it aloud for 15 minutes, score it, and write three lines: what was weak, what to check, what to read.
The incident-study habit
Once a week, read one public post-incident review of a real service. Read the review itself, not a summary.
Figure 2. From an incident review to a checklist line
How to read this: you end with a change to your own checklist, not an opinion.
Read blamelessly. The Google site reliability engineering (SRE) book says a truly blameless postmortem focuses on contributing causes "without indicting any individual or team".
Reading list: primary sources
| Source | Read it for |
|---|---|
| Anthropic, Building effective agents | Workflows versus agents; start simple |
| Lewis et al., Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks | The paper that proposed retrieval-augmented generation (RAG) |
| NIST, AI Risk Management Framework | A voluntary framework for managing AI risk |
| OWASP, Top 10 for LLM Applications | A risk list for large language model (LLM) applications |
| Google, Postmortem Culture | Reading and writing incident reviews |
What this plan does and does not give you
Expertise comes from repeated delivery: designing, shipping, watching it fail and fixing it. A course, score or plan cannot replace that, and completing this card does not make anyone an expert or guarantee a job.
The card aims at practitioner-level ability to design, build and defend an enterprise AI solution from a blank page. The 90 days deepen it through real projects.