work
projects, code, and experiments — active and archived.
colibri — 744B AI on a $500 rig
an experiment in streaming a 744B mixture-of-experts model from disk — keep ~10 GB resident in RAM, page experts in on demand, and run a frontier-scale model on commodity hardware. it runs ~1–2 tokens/sec: not fast, but a try, and proof the approach works — all on the same 2016 rig i still run today.
python · llm inference · Ryzen 7 5700G · EVGA GTX 1080 8GB · 48GB RAM · github ↗
pokeai — cooperative CPU+GPU RL
a distributed deep-RL framework that trains Pokémon Red across 256 GPU + 8 CPU instances, blending both systems' weights into each other every 5 min ("asymmetric hardware-heterogeneous cooperative RL"). runs on a 2016 GTX 1080, no cloud. still cooking — workers on Route 1.
python · cuda (sm_61) · pytorch · docker · read more →
pokemon-rl — deep RL system
end-to-end AI pipeline that learns to play Pokémon Red. Custom Game Boy emulator (~3,200 fps) + Stable Baselines 3 PPO, 64 parallel envs via shared-memory IPC, CUDA inference.
python · pytorch · cuda · stable-baselines3
rlgb — Game Boy emulator
cycle-accurate DMG emulator in C. SM83 CPU, scanline-accurate PPU, MBC1/3/5, deterministic save-states. ~3,200 fps, powers the RL training runs.
c · high-performance computing · github ↗
mc-rust — Minecraft server
Rust Minecraft server on the Valence framework. Skyblock on a void world, multiplayer, block break/place, survival, chat. Minecraft 1.20.1.
rust · valence · async networking
builtnottaught.com
this site — Rust, Axum, SQLite, Tera. Async Tokio runtime, no JS, no web fonts, single inline stylesheet. Lighthouse 100 across the board.
rust · axum · sqlite · tera · github ↗
systems & infrastructure
home network (pfSense), server administration, Docker containerization, systemd, nginx reverse proxy + HTTPS.
linux · pfsense · docker · nginx · systemd
cuda & gpu computing
GPU-accelerated experiments: CUDA kernels, parallel-processing optimization, HPC for ML training.
cuda · cudnn · pytorch
archive & experiments
old runs, dead ends, and half-finished ideas — 7z'd and shelved. not gone, just resting until i come back to them. kept as backups of what worked and what didn't.
- pokemon RL experiments — iterations with different architectures and reward shaping
- emulator variants — C, Rust, and vectorized Game Boy implementations
- training runs & checkpoints — saved sessions, model checkpoints, TensorBoard logs
- research & docs — write-ups on emulation, RL training, perf, deployment