Start here on day one. It separates the AI engineer role — building with models — from the ML researcher role that trains them. Most people waste months learning the wrong half. Do not try to complete it; use it as a map to check where you are.

The backend and DevOps roadmaps matter more than beginners expect. Almost everything I shipped at Kyndryl was an API, a container and a pipeline wrapped around a model — the model was rarely the hard part.

Free, complete, and readable without a lecturer. You need probability before anything about model uncertainty, evaluation or Bayesian methods makes sense. Thompson Sampling in my bandits project is just this maths applied.

The best deep learning book available, and every chapter is runnable code next to the derivation. Work through it with the notebooks open — reading it passively teaches you very little.

A broad sweep of classical ML before you go deep on neural networks. Useful for seeing the whole landscape quickly. My own worked notebooks for this are in the A-Z-Machine-Learning repo.

The highest-signal problem list there is. 150 problems covering every pattern you will be asked about. Do them by pattern, not at random, and redo the ones you failed a week later.

My own submissions, in public and unpolished. Posted so you can see that this is ordinary grind work — not something that comes easily to anyone.

Where you learn that real datasets are broken, imbalanced and mislabelled. Read the top-scoring public notebooks on any competition — that is a free apprenticeship in feature engineering.

Free, well-sequenced, and it covers the deployment and responsible-AI material that self-taught paths usually skip entirely. The sandbox environments mean you practise on real cloud infrastructure.

Go here once you can already code. Proper training on prompt engineering, tool use and agent design from the people building the models — this is the current frontier of the job, and most curricula have not caught up to it yet.

Write your resume in LaTeX here. It survives ATS parsing better than a Word template, it stays consistent when you tailor it per application, and you will need it for papers regardless.

Run models locally, for free, with no API key. Nearly every project on this site was built against local inference first — it removes cost anxiety from the learning process entirely, which matters more than it sounds.