A parallel, hands-on thread that runs alongside the lessons. Each build takes one idea from the course and makes it physical: you write the mechanism yourself, in numpy and the standard library, and watch it work.
Five builds are ready. Each attaches to the lesson that supplies its idea and rejoins the course straight afterwards, so a build never pulls you off the main path.
Catalogue the hardware you actually have (CPU, RAM, any GPU or boards) in a short markdown file. The baseline every later build measures against.
B1A Q-learning agent that learns to cross a 5 by 5 gridworld with a literal Q-table, no neural net. Watch a policy form from reward alone.
B2Build a minimal byte-pair-encoding tokenizer by hand: count pairs, merge the most frequent, encode and decode, and compare characters-per-token across languages and code.
B3Load pretrained word vectors as data and build the geometry by hand: cosine similarity, nearest neighbours, analogy arithmetic, and a 2D scatter.
B4Derive a 2D loss surface's gradient, hand-write the descent update, and plot the optimiser's path over contours while you sweep learning rate and momentum.
The Build Track runs to B17. Later builds unlock as the course reaches the lessons they depend on, so they are not linked yet. They are listed here so you can see the shape of the track: numpy-first implementations of the core machinery, then local deployment.