Part 2: The AI Engineer Roadmap
From your first API call to shipping production AI — the ordered curriculum that pairs with this guide.
The rest of the book is organized by topic — foundations, lifecycle, the stack, contexts (solo / startup / enterprise), patterns, decisions, career. This chapter is organized by progression. It tells you what to learn, in what order, with the why at each step and a concrete artifact to cement it.
The two views overlap deliberately. When a stage says "learn tool calling," it links into the Tool-use foundations page for depth. When the Modern Stack section says "adopt Braintrust for evals," it links into the Eval tools page for context. The roadmap is the path; the rest of the guide is the terrain.
Becoming a good AI engineer isn't one skill — it's roughly ten overlapping skills, and the order matters. Most beginners thrash because they try to build agents before they've made their RAG work, or copy LangChain code without understanding what chat.completions.create returns, or chase the model leaderboard before they've ever written a single eval. The order here is the order that actually compounds.
The four parts
| Part | What it covers | For whom |
|---|---|---|
| I. From Zero | 10 stages, ~3–6 months part-time, takes you from "I've heard of LLMs" to "I shipped a production AI feature with evals" | Working engineers new to AI |
| II. The 2026 AI Stack | The tier list: frontier vs workhorse vs cheap models, vector DBs, frameworks, eval tools, observability, gateways, plus the trends shaping next year | AI engineers picking tools |
| III. Beyond the Stack | Prompting as craft, eval mindset, retrieval quality, agent discipline, cost/latency intuition, safety — the skills no SDK teaches you | Mid-level AI engineers |
| IV. Meta-skills | How to learn AI fast when the field changes monthly, which papers to read, which communities to join, when to pivot | Anyone losing momentum |
Where to enter
- Never made an LLM API call? → Part I, Stage 0. Don't skip.
- Made a few calls, want to know what the modern stack looks like? → Part II. Use Part I's later stages as a self-check.
- Shipped AI in production, looking to level up? → Part III. The intuitions that survive every model release.
- Field is moving too fast, you're stuck? → Part IV. Being behind is the default state in AI; having strategies for it is the skill.
How the roadmap pairs with the rest of the guide
The chapters that follow this one (Foundations, Lifecycle, Stack, Solo/Startup/Enterprise, Decisions, Patterns, Career) are reference material. You don't read them straight through — you reach for them when a roadmap stage points you at them.
For example, Stage 5 — RAG links into:
- Foundations: Embeddings — for the underlying vector math
- Stack: Vector databases — for the choice matrix
- Patterns: RAG in production — for the failure modes once it leaves your laptop
- Decisions: Prompt vs RAG vs fine-tune — for when to even reach for RAG in the first place
What this roadmap is not
- Not a course on ML or deep learning. You can ship extremely useful AI without knowing how gradient descent works — and you should, before you go deeper.
- Not a tour of every framework. We pick one stack per stage; alternatives are listed in Chapter 4.
- Not about chasing the model leaderboard. The base skill — make an LLM do something useful, reliably, in production — barely changes with each new model.