AI drum transcription platform. Give it any song — it gives you a playable drum beatmap back.
Dual-model PyTorch CNN ensemble trained on 16.9M audio samples — one model on clean audio for body drums, one on Demucs-separated stems for cymbals. 7.1M parameters · 94.55% balanced accuracy · F1 0.9153 across 12 drum classes. Includes pitch-preserved tempo adjustment (50–200%) so you can slow a song down to learn it without the pitch shifting.
Stack: .NET 8 desktop client · FastAPI backend · React 18 + TypeScript · PostgreSQL · Redis · Modal GPU compute · AWS S3 · Cloudflare Pages
| 🥁 BeatSight |
AI drum transcription — trained on 16.9M samples, live in production at beatsight.io |
Python PyTorch FastAPI React .NET |
| 🎮 SSBU Adaptive AI |
Skyline plugin for Nintendo Switch — hooks Smash Bros. Ultimate's AI at the AArch64 instruction level. Persistent opponent habit learning, human-realism engine (reaction windows, execution noise), persona fitting. 92K+ lines, 546 tests. |
Rust AArch64 |
| 🎵 Mapperatorinator |
Contributor — 219M-param Whisper encoder-decoder that generates rhythm game beatmaps from mel-spectrograms. Wrote 6,000+ lines across 11 features including batch queue, audio fingerprinting, and bf16 auto-detection. |
Python PyTorch |
| 🗳️ Raptors Voting |
Full-stack election data platform processing 26,636+ records across 11 MongoDB collections — geographic choropleth visualizations, VRA ecological inference modeling, 21-stage ETL pipeline |
Spring Boot MongoDB |
