Projects — Applied Capstones
End-to-end projects that exercise the ml_first_principles
library on real tasks. Each follows the standard subproject layout
(README.md, requirements.txt, data/, notebooks/, src/, tests/, reports/),
trains in under 30 seconds with fixed seeds, uses no downloads (bundled/embedded data
only), and commits its generated report.
| Project | Exercises | Headline result |
|---|---|---|
| tabular_benchmark | Linear, tree, ensemble, distance, probabilistic models vs sklearn | Closed-form models match sklearn to 4dp; speed is the honest gap |
| char_transformer_tiny | NumPy causal transformer with manual backprop (topics 13, 16, 21) | Loss 4.06 → 2.23 in 400 steps; full finite-difference gradient check |
| digits_autoencoder | nn_core MLP + autoencoder vs PCA (topics 13, 17, 10) |
97.1% accuracy; AE beats rank-2 PCA reconstruction |
| rl_gridworld | Q-learning vs value iteration (topic 18) | Q-learning reaches the exact optimal return (9.50) |
Run any project from the repo root, e.g.:
python projects/tabular_benchmark/src/tb_benchmark.py
pytest projects # fast test suite for all projects
API friction discovered here is filed in CHANGELOG.md [Unreleased]
and drives the library's v0.2.0 scope (CONTRIBUTING.md Phase 5).