This is where I'm learning everything related to deep learning — out in the open, as I go.
Instead of picking one topic and grinding through it linearly, I run multiple learning roadmaps at once and jump back and forth between them. I try to tackle the hardest topics first — the point is to move fast, follow curiosity, and let progress on one topic inform the others, rather than getting stuck waiting to "finish" a track before starting the next.
The goal isn't polished courseware. It's learning as fast as possible through real experiments and real code — this repo is the trail of that process, not a finished product.
Every topic follows the same layout:
<topic>/
docs/
roadmap/
00-overview.md
01-...md
02-...md
...
Code and experiments live alongside the docs, in the same topic directory, as each roadmap is worked through.
Use the writing-learning-roadmap Claude skill to generate a new topic's roadmap.
Use the verifying-exercise-solutions Claude skill to check an exercise file against its roadmap doc.
Dependencies (PyTorch, NumPy, etc.) are managed with uv in a single shared environment at the repo root. From anywhere inside the repo, run a script with:
uv run python path/to/script.py
uv automatically finds the root pyproject.toml and uses the shared .venv, no manual activation needed. To add a new dependency, run uv add <package> from the repo root.
MIT — see LICENSE.