Path to AI Engineer
Learn AI engineering from scratch - seven free resources, in order.
The exact path if you were starting this year. No paid courses, no signups, no shortcuts. Open the first link, finish it, then open the next.
7
Free resources
$0
Total cost
~6 mo
Honest timeline
AI engineering isn't one subject. It's seven, stacked.
Most people drown trying to learn AI because they treat it as a single thing to pick up. It isn't. It's a stack: language, theory, math, models, retrieval, building, and orchestration; and each layer only makes sense once the one below is solid.
The good news is the best resource for each layer is free, and most of them are taught by the people who built the field. Andrew Ng. Andrej Karpathy. The LangChain team. You don't need to find them. They're below, in order.
The hard part isn't access. It's discipline. The temptation is to skip Python and jump to building agents because agents look like the fun part. Resist it. The people who stick at this finish each rung before climbing the next, and finish all seven inside six months.
The rule: open the first link today. Don't open the second until you've actually finished the first. The path only works if you walk it in order.
STAGE 01 · LANGUAGE
Python first. Not negotiable.
The language you'll write every other layer in.
You can't learn AI without Python. Every framework, every notebook, every model API. If you're starting from zero or your Python is rusty, this is where you spend the first six to eight weeks; no detours.
RESOURCE · COURSE · FREE
How to use it: don't skip the problem sets. The lectures teach syntax, the problem sets teach you to think
STAGE 02 · THEORY
Machine learning fundamentals.
The floor underneath every model you'll ever use.
Before LLMs, before transformers, before agents, there's the basic shape of machine learning. Supervised vs unsupervised, training and loss, overfitting, evaluation. Skip this and the rest of the stack feels like magic instead of engineering.
Andrew Ng's specialization is still the gold standard for this clean, rigorous, free to audit, and it assumes you're starting from zero.
RESOURCE · COURSE SPEACIALIZATION · FREE TO AUDIT
How to use it: watch each lecture, do the lab. If the math in a lecture loses you, pause and open the next resource; that's exactly what it's for.
STAGE 03 · MATH
The math underneath the models.
A reference you keep open the entire time, not a course you finish.
Linear algebra, calculus, probability, statistics. You don't need a degree in any of them, you need enough to read a paper without bouncing off the equations and to understand what's happening when a model trains.
The right way to learn this is as an ongoing thing, alongside the ML specialization. Bookmark the resource below and open it every time a term lands you stuck.
RESOURCE · REFERENCE · FREE
How to use it: don't skip the problem sets. The lectures teach syntax, the problem sets teach you to think
STAGE 04 · LLMS
How language models actually work.
Watch it twice. The first time for vibe, the second with a notebook open.
This is where most self-taught AI engineers separate from the pack. The people who can use an LLM are everywhere, the people who understand what it's doing under the hood are rare. Karpathy's playlist is the best free explanation of that, anywhere.
He builds a language model from scratch, in a notebook, line by line. By the end you understand tokens, attention, training, fine-tuning; not as words, but as code you've typed.
RESOURCE · YOUTUBE PLAYLIST · FREE
How to use it: watch each video once at normal speed for the shape of the idea. Then watch it again with the notebook open, and type along. The second pass is where the understanding lands.
STAGE 05 · RETRIEVAL
RAG - the skill the market is paying for right now.
How you make an LLM actually useful on your company's data.
A plain LLM is impressive but generic. The moment a company wants to use one against its own docs, contracts, tickets, or knowledge base, you need retrieval-augmented generation. It's the single most in-demand applied skill in the field right now.
The freeCodeCamp tutorial below, taught by a LangChain engineer, is the most thorough free introduction to it on the internet.
RESOURCE · YOUTUBE TUTORIAL · FREE
How to use it: build it as you watch. By the end, you should have a working RAG pipeline against a small dataset of your own: your notes, your PDFs, anything. Don't move on until you do.
STAGE 06 · BUILD
Now build things. Real things.
Theory is half the work. The other half is shipping.
At this point you've read, watched, typed along. The next thing is to build your own projects without a tutorial guiding you. This is the part where most people stall, staring at a blank repo, not sure what to build.
The repository below solves that problem. Seventy-plus end-to-end tutorials covering LLMs, RAG, agents, MCP, LLMOps, and evaluations; each one a small project you can fork, ship, and put on a resume.
RESOURCE · GITHUB REPOSITORY · FREE
How to use it: pick one project a week, build it, then change one thing about it: different data, different model, different output shape. The changes are where you actually learn.
STAGE 07 · AGENTS
Agents. Where the field is going.
Models that take actions, not just answer questions.
Everything before this stage is about getting a model to give you a good answer. Agents are about getting a model to do a task: call tools, make decisions, hand off work, recover from errors. The next two years of AI engineering jobs are mostly here.
LangChain is the most widely-adopted framework for this, and they teach it for free through LangChain Academy. Pair the course with the official agents documentation.
RESOURCE · COURSE + DOCS · FREE
How to use it: do the course, then rebuild one of your Stage 06 projects as an agent. A chatbot that answers questions is a Stage 05 project; one that books a meeting on your calendar is an agent.
The seven resources at a glance
How to start.
Open the CS50P link today. Not bookmark, open. Watch the first lecture this evening, do the first problem set this weekend. That's the whole start.
From there, the path is mechanical. Finish each resource before you open the next. Don't read three Twitter threads about agents while you're still on Python, that's how six months becomes two years.
Build small things from week one. Even in Stage 01, write a script that does something for you. A folder organizer. A daily log. The earlier you treat code as something you ship, not something you study, the faster every later stage goes.
Six months from now, the people who finished this path won't have a different degree. They'll have a different relationship with their own time, and the entire AI job market will be open to them.