AI Engineering Self-Study¶
A 13-week builder course: you build one product, an anti-scam message checker, and learn each skill at the moment the product needs it. The AI work (agents, RAG, evals, fraud data) comes first. Engineering and DevOps tools arrive only when something breaks or needs shipping.
The course started from the modules of NTU's SCTP AI Engineering programme, recalibrated: less web development, more agents, evals and fraud data at scale.
The 13 weeks¶
| Week | Starts | Focus | What you can show at the end |
|---|---|---|---|
| 1 | Mon 14 Dec 2026 | Checker v0 | An LLM checker that beats a keyword baseline, with the numbers |
| 2 | Mon 21 Dec | Agent loop by hand | A hand-written tool loop with tested tools |
| 3 | Mon 28 Dec | State, MCP, a framework | Replayable runs, a checker step, one MCP server |
| 4 | Mon 4 Jan 2027 | RAG | Retrieval over known scam patterns, with recall numbers |
| 5 | Mon 11 Jan | Evals I: error analysis | 100 labelled messages and a ranked list of failure types |
| 6 | Mon 18 Jan | Fraud data at scale | A PySpark pipeline and graph features over scam reports |
| 7 | Mon 25 Jan | Evals II: judges and CI | An LLM judge and a CI quality gate |
| 8 | Mon 1 Feb | Security | A red-team report and guardrails |
| 9 | Mon 8 Feb | Real backend | A FastAPI and Postgres service deployed on Cloud Run, plus your first public link |
| 10 | Mon 15 Feb | Front end and users | A streaming chat screen, tested on 10 people |
| 11 | Mon 22 Feb | Production | Cost and latency report, an AWS deploy |
| 12 | Mon 1 Mar | Ship | A public case study |
| 13 | Mon 8 Mar | Buffer | Repeat whichever week didn't click |
The dates assume a start on Monday 14 December 2026. Christmas (Fri 25 Dec), New Year's Day (Fri 1 Jan) and Chinese New Year (Sat 6 – Sun 7 Feb 2027, with Mon 8 Feb as the holiday in lieu, the first day of week 9) fall inside the course. Weeks 2 and 3 are planned as four-day weeks: week 2's demo moves to Thursday afternoon, and week 3's demo moves to Monday 4 January, so week 4 starts after lunch that day. For other holidays, push the week's last task into the next Monday rather than working the weekend. Finishing in late March instead of mid-March is fine.
How Claude fits in¶
- Weeks 1–8: you call Claude from Python through the Claude Code command line, using your subscription. No API key, no extra cost. Calling Claude from Python sets this up once and every week reuses it.
- Week 9 onwards: the checker goes live for other people, which needs an Anthropic API key with a spend cap. You swap one function and nothing else changes.
This is also why the first public link arrives in week 9: a subscription login can't power an app that strangers use. Weeks 1–8 run on your own laptop.
Target roles¶
AI engineer on small teams, AI evaluation roles, and anti-fraud and anti-scam engineering teams. Week 6 exists for the last group. The filter for any role: success is settled by tests and working software, not negotiated between business teams.
Apply in waves as each piece of evidence lands:
- after week 3: small-team AI engineer roles (you have a working product and an agent you built by hand)
- after week 5: evaluation roles (you have labelled data and error analysis)
- after week 6: anti-fraud and anti-scam roles (you have PySpark and graph work on fraud data)
Daily rhythm¶
Treat it like a full-time class: same place, same hours, weekdays only, about 25 hours a week. Studying at a fixed spot away from home, such as a public library, is what replaces the classroom.
| Time | What |
|---|---|
| 9:00–9:15 | Write today's one goal on paper; check yesterday's commit |
| 9:15–10:30 | Learn: the day's course, docs or video |
| 10:30–13:00 | Build: apply it to the week's deliverable |
| 13:00–14:00 | Lunch, walk |
| 14:00–15:30 | Build or fix; finish with a commit to GitHub |
| After 15:30 | Gym, rest, family; no study |
Weekly:
- Monday: copy the week's checklist into your project README and plan the week
- Friday 9:00–12:00: demo day. Record a 3-minute screen video of what works, write five lines on what you learned, and send both to Claude using the Friday check-in
- Friday afternoon: from week 9, the optional cloud side track
- Weekends: off
Rules: every weekday ends with a commit. If a week slips, cut scope, not the deliverable. A module that hasn't clicked gets one repeat week; week 13 is the buffer for that.
How each lesson works¶
Every week page has the same shape:
- Why this week: what the product needs and why
- Concepts: the ideas in plain language, before any code
- Monday to Friday: what to read in the morning and what to build after, with code to start from
- Done when: checks that tell you the day is finished
- Stuck?: hints hidden behind a click, so you try first
- Friday: the deliverable, the demo and the check-in
You build in your own project repository (scam-checker), which week 1 sets up. This site is only the course.
Engineering and DevOps, only when needed¶
| Week | What forces it | What you add |
|---|---|---|
| 1 | You need a repo you won't regret | git, uv, ruff |
| 2 | Tool functions start breaking in ways you can't see | pytest with the LLM and network mocked; mypy |
| 4 | RAG needs a database | Postgres with pgvector under docker compose |
| 7 | A prompt change must not quietly lower scam recall | GitHub Actions running tests and the eval suite |
| 8 | Attackers go after dependencies and secrets too | Dependency, image and secrets scanning in CI |
| 9 | Real users need a real service | FastAPI, Alembic, a Dockerfile, Cloud Run, Secret Manager, deploy on merge |
| 11 | You need to know cost, speed and failures | Tracing, JSON logs, a health check, AWS |
If something hurts earlier than this, pull it forward. Keep DevOps with Docker and The Missing Semester as references to dip into, not courses to work through.
In-person anchors¶
The NUS-ISS Building Autonomous AI Agents course fits weeks 2–3, and optionally Deploying Safe and Secure AI Agents fits week 8. Check dates early; they don't run every month.
What self-study misses, and how to cover it¶
| A programme gives you | Self-study substitute |
|---|---|
| Instructors who review your code | Ask Claude or a study partner to review each week's deliverable before you move on |
| A cohort and fixed class hours | The daily rhythm above, an in-person course in weeks 2–3, and the Friday check-ins |
| Assessed capstone and certificate | A deployed product, a public case study with numbers, and the repo |
| Career consultation | Your network, and volunteering with an anti-scam group such as CheckMate |
Costs¶
| Item | Cost |
|---|---|
| All main courses and docs | Free |
| Claude, weeks 1–8 | Included in your Claude subscription |
| Claude API and cloud hosting, weeks 9–12 | Roughly S$30–80 (an estimate); set a billing alert and spend cap |
| Optional NUS-ISS in-person courses | About S$932 each after funding; SkillsFuture Credit can cover them |
| NTU programme, for comparison | About S$9,350 after funding |