Learn ML by writing `loss.backward()` yourself, in NumPy

What it does

TrenTorch is an educational project with two faces: a CLI curriculum of 20 modules where you build a deep learning framework in raw NumPy — tensors, autograd, optimizers, convolutions, attention, all the way to a capstone — and a web app with 350+ coding questions across 12 topic areas, from linear algebra to production ML serving. The browser version runs Python via Pyodide and grades your code instantly, no install required. There is no import torch anywhere in the curriculum; you know what loss.backward() does because you wrote it.

The interesting bit

The CLI curriculum is a rebuild and extension of TinyTorch (Harvard CS249r), and as you progress you unlock recreations of landmark results — the 1958 perceptron, the 1969 XOR crisis, backprop in 1986, CNNs in 1998, transformers in 2017 — all running on the framework you built yourself. It’s a LeetCode-shaped delivery vehicle for ML fundamentals, aimed partly at interview prep.

Key highlights

• Two delivery modes: browser (SvelteKit + Pyodide, instant grading, GitHub/Google/email sign-in) and terminal/Jupyter via the tren CLI
• Coverage spans math foundations, classical ML, deep learning core, transformers/LLMs, inference (KV cache, quantization, batching), systems performance, RL, and production ML
• Every question ships a problem statement, theory section, starter file, tests, and a reference solution
• A Problem of the Day keeps the web side fresh; CLI progress is saved locally under user_data/
• 18 contributors, CI-validated, CodeFactor A+

Caveats

• PolyForm Noncommercial license — fine for learning, not for building a product on top of it
• The web app’s local development requires your own Supabase keys; the hosted version is the frictionless path
• The CLI curriculum is derived from TinyTorch rather than wholly original, which the authors state plainly

Verdict

If you’ve ever imported a framework and felt uneasy about not knowing what’s underneath, this is the antidote — especially strong for ML interview prep. If you just want to train models rather than understand them, keep importing.