A 16-week path · 96 notes · 5 projects

AI Engineer Notes

From “what is a vector?” to designing LLM systems in an interview. Every idea is explained simply first, then precisely, then in words you can say out loud, and then practised until it sticks.

These notes assume no AI background and no more math than school algebra. Each one starts with an everyday explanation, shows the idea as a picture, then gives the exact technical terms, the math and the code. It ends with a model interview answer, practice questions and a revision card.

Start with the roadmap to see what to study each week, use the practice page for the drills, and check the projects you will build alongside. 8 of 96 notes are written so far. Reference pages sit beside the notes: the glossary, one plain sentence per technical word, and three code references, for NumPy, scikit-learn and Python tools, one entry per function the code uses.

How these notes are written

Good explanations follow a few rules that teachers have known for a long time, and every note here is held to them. They are listed so you can hold the notes to them too: if a note breaks one, that is a bug.

1. Nothing is used before it is taught. The notes are read in a fixed order, and a note may only use words that an earlier note explained. When one must mention a later idea, the word is dotted-underlined and links to the glossary, with its one-line meaning on hover.
2. Concrete before abstract. Every idea starts with an everyday situation and small numbers (three houses, a coin, a recipe), then gets its picture, then its name, then its symbol. Experts tend to start with the symbol because it feels clear to them; it is not clear to a beginner, and we never do it.
3. Plain words first in every section, not just the introduction. Each section opens with a “plain words” box that a bright fourteen-year-old could follow, and only then uses the exact terms.
4. Ideas return in spirals. You meet the log loss as a forecaster’s score in week 1, as a formula in week 2, as maximum likelihood in week 3 and as cross-entropy in week 4. Each visit adds one layer.
5. Worked examples, then your own. Every note works its numbers in full before asking you to, then gives about a dozen questions and sixteen problems, hand and code, with hidden solutions.
6. Code uses only what has been explained. The code assumes plain Python and nothing else. Every library function and every less common bit of Python is introduced in a “New in this code” box right above the first block that uses it, in one plain sentence. Every name inside a code block is also a link to its entry on the NumPy, scikit-learn or Python tools reference page. Code is written one step per line, and every output shown is real.
7. Only what an AI interview needs. Each topic earns its place by one of two tests: interviews ask it, or a later note needs it. The rest is left out on purpose and named in a “what you can skip” note.

The nine tracks