Skip to main content

How to Use This Roadmap

If you're new to AI engineering​

  1. Read Part I top-to-bottom, one stage at a time. Don't skip. Each stage builds on the previous; if you skip from "first API call" to "agent," you'll spend a month debugging a loop that's broken in three places and you can't tell which.
  2. Build the artifact at the end of each stage before moving on. The teal Project callouts. Reading without building is the AI-engineering equivalent of watching swim videos and calling yourself a swimmer.
  3. Use the linked canonical resources for depth. Each stage links out to free, gold-standard learning material (OpenAI cookbook, Anthropic prompting guide, the actual papers when they're readable) and into the Foundations and Stack chapters of this guide. The roadmap gives you the path; those give you the practice hours.
  4. Don't rush. Stages 0–9 are 3–6 months of part-time work depending on your background. There's no shortcut — and the people who claim to have done it in two weekends usually have a portfolio of toy demos with no evals.
  5. When stuck, read Part IV. "How to learn fast" and "When to pivot" exist because the field is moving faster than you can keep up — having strategies for that is the skill.

If you can already make an API call and ship a prompt​

Skip the first three stages, but do Stage 6 (evals) even if you think you know it. Most "I've been doing AI for six months" engineers have never written a real eval set — and it's the load-bearing skill that distinguishes hobby from production. Then jump to Part II — The 2026 AI Stack.

If you're already shipping AI in production​

Part III — Beyond the Stack is where you live. Prompting as craft, eval discipline, retrieval quality at scale, agent reliability, cost/latency intuition, safety. None of which the SDKs teach you, and all of which become the constraint on production AI as soon as you scale past a single customer.

A note on AI-assisted learning of AI

ChatGPT and Claude can teach you AI engineering. They're also exactly the wrong teacher for half the lessons — they'll generate a LangChain snippet that "works" and skip the inspection step that's the whole point of Stage 1. Use AI to explain what a piece of code is doing ("walk me through what response.choices[0] contains") rather than to do the stage for you. The latter is how you get six months into AI engineering without understanding the message-format the API actually returns.

The legend​

Pages in this chapter use a few recurring callouts:

  • Stage — a milestone in the curriculum. Has a time budget and an artifact you should have built by the end.
  • Tier 1 / 2 / 3 — opinionated picks. Tier 1 = adopt now. Tier 2 = worth knowing. Tier 3 = skip or defer.
  • In plain English — the explanation for someone who's never heard of the concept.
  • Project — what you should build before clicking next.
  • Common mistakes — where people predictably trip up.
  • Page checkpoint — three quick questions to make sure the page stuck.

The cardinal rule​

No new stage until the previous stage's artifact runs end-to-end with at least one eval case passing.

This is the single rule that separates engineers who become senior AI engineers in a year from engineers who are still demoing the same toy chatbot after two. Build it. Measure it. Then move on.

→ Start with Stage 0 — Setup