TrenTorch
A from-scratch machine learning framework built in NumPy, paired with hundreds of browser-graded coding questions, for learning how ML and inference internals actually work instead of only calling an existing library.
TrenTorch teaches machine learning by having the learner rebuild it: a NumPy-only framework whose operations trace back to raw NumPy rather than importing PyTorch, matching the architecture of production frameworks so a concept like loss.backward() is understood because the learner implemented it. Its CLI curriculum extends a Harvard CS249r (TinyTorch) curriculum, rebuilt in TrenTorch's own style, while its browser version is built independently.
Two parallel tracks cover the same ground: a browser track with more than 350 coding questions across 12 topic areas, graded instantly through Python running in-browser via Pyodide, with progress saved to an account; and a terminal/Jupyter CLI track with 20 modules that build a framework from tensors up to a capstone project, with progress saved locally. The web app is a SvelteKit project using Supabase for sign-in.