Timeline & Suggested Order
The suggested order
Don't read the parts strictly 1 → 4. Use them as overlapping tracks:
- Always-on track — Part IV (Meta-skills). Read How to learn AI fast and When to pivot before you start Stage 1. They change how everything after them sticks — and they're cheap to read.
- Primary track — Part I, in order. Stages 0 → 9, no skipping. This is 3–6 months part-time.
- Lookup track — the rest of this guide. When Stage n points at Foundations or Stack, follow the link, read the page, return.
- Once you're shipping — Part II. As you complete Stages 6+, sample the trends and tier-1 picks against your own projects.
- Once you're mid-level — Part III. Roughly when "did the API call return JSON" stops being the hard question and "is this RAG actually good" becomes the hard one.
How long does all this take? — honest version
The marketing copy says "build AI agents in a weekend." Most of those weekend agents have no evals, no cost caps, no observability, and would catch fire the second a real user hit them. Honest engineering takes longer.
| Background | To Stage 5 (first RAG) | To Stage 9 (shipped + evaled) |
|---|---|---|
| Total beginner, part-time (10 hrs/week) | 1.5–3 months | 4–6 months |
| Total beginner, full-time (35+ hrs/week) | 3–5 weeks | 2–3 months |
| Working web/backend dev, part-time | 3–5 weeks | 2–4 months |
| Working ML/data scientist, part-time | 1–2 weeks (you know the math; you don't know the productionization) | 1–2 months |
| Working AI engineer already shipping | Days (use Part I as a self-check) | Not the target — go to Part II / III |
These are honest numbers. Plan for the upper end. If you hit the lower end, great.
How many hours per week, sustained, with builds at the end of each stage. Ten hours every week for six months beats forty hours for one month followed by burnout, every time. The dropouts in AI engineering are almost always burnout-driven or motivation-driven, not talent-driven — the field rewards the people still showing up in month four.
What to skip if you're in a hurry
If you absolutely must compress to ~4 weeks, here's the minimum viable path. You'll have gaps; you'll know what they are.
| Stage | Keep / skip | Why |
|---|---|---|
| Stage 0 — Setup | Keep | Two hours; saves you two weeks of debug |
| Stage 1 — First call | Keep | One afternoon; the mental model is the whole point |
| Stage 2 — Chatbot | Skip if you don't ship chat | Pure UI work; reusable elsewhere |
| Stage 3 — Structured output | Keep | Highest-leverage pattern in the roadmap |
| Stage 4 — Tools | Keep | Half of all AI features are this |
| Stage 5 — RAG | Keep | Half of all AI features are this |
| Stage 6 — Evals | Keep, non-negotiable | Without it you're flying blind |
| Stage 7 — Observability | Compress to: "log every call to one table" | Full hosted observability can wait |
| Stage 8 — Agent | Skip until you need it | Most production AI is not agents |
| Stage 9 — Ship | Keep | The lessons are in the deployment friction |
That's seven stages instead of ten, ~4 weeks at full-time pace. Note: the skipped stages are also the ones least useful to "demo and look smart" — they're the ones you grind out for production reliability.
What to skip if you're already a web developer
You already know hosting, auth, rate limiting, observability concepts. Compress Stages 0, 2, 9 — most of the operational work transfers directly. The new material is Stages 3, 4, 5, 6 — and especially Stage 6, because writing evals is a category of testing you haven't done before.
What to skip if you're already an ML engineer
You know about embeddings, vector math, fine-tuning, training loops. Compress Stages 3, 5 — the mechanics are familiar. The new material is Stages 4 (tool calling as a production pattern), 6 (evals as a product discipline, not a research benchmark), 7, 9 — the productionization work that academic ML doesn't cover.
A note on prerequisites
This roadmap assumes you can write a small program in some language and read English documentation. It does not assume:
- Math beyond high-school algebra (vectors and cosine similarity come up; both are learnable on demand)
- Any prior ML experience
- A CS degree (helpful but absolutely not required)
- Familiarity with cloud infra (you'll meet just enough at Stage 9)
If you're worried about whether you can do this: you can. The variable that matters is sustained hours, not innate aptitude.