I kept running into the same wall. Every time I wanted to learn a new piece of AI — say, how LoRA actually saves memory, or what "temperature" does inside softmax — I'd end up with eight browser tabs open: a blog post that hand-waved the math, a YouTube video I'd have to scrub through, a Stack Overflow answer from 2019, and a paper I couldn't finish. Nothing connected. And nothing remembered what I already knew. So I built the thing I wanted: one place where every idea is tied to its formula, you can type your own numbers in and watch it work, and the system actually tracks what you're weak at. The problem, stated plainly. Most free AI courses are content libraries. You watch, you nod, you forget. They don't measure whether you learned anything, they don't know what you already know, and they don't bring things back at the right time. Paid platforms do some of this — but they gate it behind subscriptions and they still don't show you the statistics. What's actually different here. I'm a data scientist, so I leaned into that. Every quiz answer feeds four things at once: An Item Response Theory ability estimate (θ) per subject, with a 95% confidence interval from the Fisher information. The same family of models the SAT and GRE use. A Wilson score interval on your accuracy — because the normal approximation lies when n is small. A calibration curve and Brier score — you tell us how sure you are before each answer, and we plot whether your confidence matches reality. Almost nobody ships this. SM-2 spaced repetition — so missed questions come back in a day, mastered ones fade to a week, then a month. The dashboard shows you all of this. On purpose. If you don't trust the numbers, you shouldn't trust the course. What I built on top. 66 interactive topics from vectors and gradients, all the way up to transformer architecture, RAG pipelines, agent security and LLM evaluation. 46 live calculators — you type numbers, the formula runs, you see the answer. Not screenshots of formulas. Real arithmetic. Feynman teach-back — after each topic, explain it in your own words. You get graded by keyword coverage, or by your own OpenAI / Anthropic / Groq key if you've set one. A prerequisite knowledge graph of all 66 topics, colored by your mastery, with the frontier — what you're ready to learn next — highlighted in orange. A browser code runner with Python (via Pyodide/WASM) and SQL (via sql.js). Nothing uploaded. Nothing installed. A content authoring tool so you can add your own topics and export them as JS. What it deliberately is NOT. No accounts. No email capture. No tracking pixels. No ads. Everything lives in your browser's localStorage. When you opt in to share an anonymous θ with the community cohort, we send exactly four numbers — subject name, ability, correct count, total count — and nothing else. The methodology page spells out every model, every assumption, and every limitation. I'm not trying to build a business here. I'm trying to build the thing I wish existed when I started. What I'd love from you: If you try it, tell me one topic that felt weak or one formula that didn't land. I'll fix it. If you're a teacher or a team lead and want a self-hosted version with your own content, comment below — I've been thinking about that. Thanks for reading. Ask me anything. 🙏

MathToModels

Learn AI engineering from the math up — interactive & free

Learn AI engineering from the math up — interactive & free