ML from scratch, with moving pictures to prove it

What it does

This repo is the build system for a collection of notebooks that implement neural networks, logistic regression, PCA, k-means, and gradient descent from first principles using NumPy. Each algorithm lives in its own repository; this one fetches them via shell script, stitches them into a Jupyter Book, and publishes a website. The output is animated: you watch loss surfaces flatten and decision boundaries shift as weights converge.

The interesting bit

The project treats visualization as proof, not decoration. Marimo interactive notebooks let you tweak weights and see loss functions respond in real time—useful for anyone who suspects backprop is just matrix witchcraft until they see the gradients flow.

Key highlights

• Six algorithms rebuilt from scratch: no sklearn black boxes
• Marimo interactive notebooks for live parameter manipulation
• LaTeX-derived math explanations alongside working code
• Automated site builds via GitHub Actions on every commit
• Also exports to EPUB if you prefer your gradients in e-ink

Caveats

• The actual notebooks aren’t in this repo; you need to run a shell script to pull them from scattered repositories, which feels like buying a cookbook where the recipes are mailed separately
• The README’s “Usage” section is mostly Docker incantations and build steps rather than guidance on the math itself

Verdict

Good for students or self-taught developers who need to see the plumbing before trusting the API. Skip it if you want production-ready implementations or are already comfortable reading Bishop cover-to-cover without visual aids.