Become an AI Engineer — The Practical Guide

The Role

What an AI Engineer actually does

An AI Engineer builds products on top of foundation models — GPT, Claude, Gemini, Llama — rather than training models from scratch. The job is software engineering with a new toolbox: calling model APIs, grounding them in your data, giving them tools, and proving the whole thing works before shipping it.

Build with models

Design prompts, structured outputs, and function calls that turn a raw model into a reliable product feature.

Ground them in data

Connect models to your documents and databases with retrieval (RAG) so answers are accurate and cited.

Give them agency

Build agents that use tools, browse, write code, and complete multi-step tasks — safely and observably.

Prove it works

Write evals that measure quality, catch regressions, and separate a demo from a production system.

AI Engineer vs. the neighbouring roles

Role	Core question	Typical work	Math depth needed
AI Engineer	How do I build a product with existing models?	LLM APIs, RAG, agents, evals, deployment	Low–medium — engineering matters more
ML Engineer	How do I train & serve custom models?	Training pipelines, feature stores, model serving, MLOps	Medium–high
Data Scientist	What does the data tell the business?	Analysis, experimentation, statistical modelling	High (statistics)

Why the role exists now: foundation models made state-of-the-art AI available behind an API call. The scarce skill is no longer training models — it's building, grounding, and evaluating products on top of them. That's the AI Engineer's job.

Skill Map

The seven skills that matter

Everything on the roadmap below builds one of these. If a tutorial doesn't feed into one of them, skip it.

Python + APIs

Solid Python, HTTP, JSON, virtual envs, git. The unglamorous 80% of the job.

How LLMs work

Tokens, embeddings, attention, context windows — deep enough to predict failure modes.

Prompting & structured output

System prompts, few-shot examples, JSON mode, function/tool calling.

RAG

Chunking, embeddings, hybrid search, reranking, citations. The #1 asked-for skill in job posts.

Agents

Tool use, MCP, planning loops, multi-agent patterns — and when a plain pipeline beats an agent.

Evals

Golden datasets, LLM-as-judge, regression tests. The skill that separates seniors from tutorial-followers.

Shipping

FastAPI, Docker, streaming, cost/latency budgets, monitoring. A model in a notebook isn't a product.

The Roadmap

Five phases, each ending in something you built

Go phase by phase and finish each build before moving on. Every phase maps to a section on the resources page. Rough pace: 2–4 weeks per phase at ~8–10 hrs/week.

Foundations

Get fluent in Python and understand what an LLM actually is — no math degree required.

Python: functions, classes, typing, venvs, requests/httpx
Watch how neural nets and transformers work (3Blue1Brown, Karpathy)
Understand tokens, embeddings, context windows, temperature

Build: a CLI script that calls an LLM API, streams the reply, and counts tokens & cost per request.

LLM APIs & Prompting

Turn a raw model into predictable, structured behaviour.

System prompts, few-shot prompting, chain-of-thought
Structured outputs (JSON schema), function/tool calling
Handling failures: retries, timeouts, fallbacks, rate limits

Build: a document extractor — PDF/email in, validated JSON out, with tests proving the schema always holds.

RAG — Retrieval-Augmented Generation

Ground the model in your own data so answers are accurate and cited.

Embeddings & vector search; chunking strategies that don't destroy meaning
Hybrid search (keyword + vector) and reranking
Grounded answers with inline citations; when RAG fails and why

Build: a Q&A app over a real document set (docs, papers, wiki) with citations — not a 20-line framework demo you can't explain.

Agents

Let models take actions: call tools, run code, complete multi-step tasks.

The agent loop: reason → act → observe; tool schemas and MCP
Multi-agent patterns, guardrails, human-in-the-loop
Knowing when not to use an agent (most of the time)

Build: an agent with 3–4 real tools (search, calculator, your RAG app, code runner) that completes tasks end-to-end with logging.

Evals & Shipping

Prove it works, then put it on the internet. This phase is what gets you hired.

Golden datasets, LLM-as-judge, pass/fail regression suites
FastAPI + Docker deployment, streaming responses, monitoring
Cost & latency budgets; caching; graceful degradation

Build: add an eval suite to your RAG app, deploy it publicly, and write a README that shows your quality numbers.

Ask the Guide

Questions? Ask the AI.

This box is itself an example of what you'll build in Phases 2–3: an LLM grounded in this site's content, running serverless on Cloudflare Workers AI.

The Payoff

How to actually get hired

Certificates don't get AI engineering jobs — evidence does. Three things make up the evidence:

2–3 public projects

Deployed, documented, with real data. One great RAG app with evals beats ten tutorial clones.

READMEs that teach

Explain the decisions: why this chunking, why these evals, what failed. That's what interviewers read.

Write as you learn

Short posts on what you built and measured. Public writing compounds into interviews.

The one-line strategy: don't try to learn everything — build the RAG → Agents → Evals → Ship spine end-to-end and make it public. That combination is what separates “ran a notebook” from “hired”.

Start learning — curated resources →