AI Engineering Roadmap
A sequence for frontend engineers learning AI engineering: LLM APIs, RAG, agents, evals, then production. Checkpoints you can build.
- roadmap
- ai-engineering
- career
- learning-path
This path assumes you already ship JS/TS. You do not need a year of machine learning first. You need a model of tokens, a retrieval loop you can eval, and the same production habits you use for payments.
Full notes live under AI engineering. Use this page as the order.
Stage 0 — Prerequisites
- TypeScript,
fetch, HTTP status codes - A backend-for-frontend instinct: secrets stay off the client
- Comfort reading official API docs
Checkpoint: a Route Handler that calls a model with a server-side key and streams tokens to the browser.
Stage 1 — LLM literacy
Goals: treat the model as a sampler with a context window.
- What is AI engineering?
- LLMs for frontend engineers
- Tokens and context windows
- Chat messages and system prompts
- Temperature, top-p, and sampling
- Structured outputs
Checkpoint: extract a typed object from messy user text with a schema. Fail closed on parse errors.
Stage 2 — Prompts as code
Checkpoint: version a prompt, log it, and abort in-flight generations from the composer.
Stage 3 — Retrieval (RAG)
- RAG mental model
- Chunking
- Embeddings and similarity search
- Hybrid search
- Citations
- When RAG is the wrong tool
Checkpoint: a docs chatbot that cites real URLs from your corpus and refuses when retrieval is empty. Walkthrough: building a docs RAG chatbot.
Stage 4 — Tools and agents
- What are AI agents?
- Think, act, observe
- Function calling
- Tool schema design
- MCP
- Human in the loop
- Failure modes
Checkpoint: one read tool and one side-effect tool. The side-effect waits for approval. Max steps are enforced in code.
Stage 5 — Quality and safety
Checkpoint: a 30-case golden set in CI. A forbidden document never appears in the prompt.
Stage 6 — Production
Checkpoint: p95 TTFT and cost per successful answer on a dashboard you actually open.
Stage 7 — Interviews
Checkpoint: talk through a refund bot on a whiteboard: tools, auth, evals, cost, and the approval UI.
What to skip (for now)
- Training a base model
- Fine-tuning before you have an eval the prompt cannot move
- Multi-agent crews before a single tool loop is boring and reliable
Do you need ML first? if that question is blocking you.