I make pathology slides talk, and big models run on small computers.
🔬 Postdoctoral Research Fellow @ CHIP — Boston Children's Hospital / Harvard Medical School ⚙️ CTO @ Tropiflo 📚 Google Scholar
Predicting cancer outcomes from whole-slide images: foundation models, multiple-instance learning, and multi-center survival analysis.
- PATH-ORACLE-INFERENCE — weights, cutoffs & inference for our LUAD progression models (TITAN embeddings ± a 46-gene RNA signature → risk)
- SlideCrush — patch & embed WSIs with every histopathology ViT foundation model in one package
- onco_run — model-agnostic WSI inference runner in a single Docker image; collaborators just
./run.sh - KAIROS & PRELUDE — international consortia: platinum response in ovarian cancer, and early-stage lung adenocarcinoma relapse
- 🐦 colibri — hacking on a pure-C MoE engine that streams experts from disk. Current experiment: a ~1T-parameter model running off a spinning hard drive
- 🖥️ sovereign-mesh — interactive simulator: wire your office's idle Macs into one machine big enough for frontier open-weight models
- 🎙️ voice-claude — "Hey Jarvis" → talk to your coding agent through a Bluetooth headset, Whisper running on the Apple Silicon GPU
- ☁️ runway — see your cloud credits burning down without leaving your coding agent
- 📈 polymarket-kalshi-btc-arbitrage-bot — does what it says on the tin
Proof I'll build anything once: a speed-dating site named DinosaurDating, a fridge app that nags you about expiring food, a pandemic-era CNN that livestream-detects you touching your face, and a GNN for pigeons.
Currently obsessed with: how much frontier-model inference you can squeeze out of hardware nobody wanted. Ask me about R(3,3,3,3). 🎨