build-small-hackathon/the-committee
🐱 The Committee
Two house cats intercepted your webcam. They understand exactly what you're doing. They are judging you for it.
Point your camera at yourself and Mochi (a manic kitten who must attack everything) and Loaf (an enormous, withering elder who lives for food) watch through the window and react like cats — demanding food, plotting, holding grudges, pouncing at the glass. Built for the Hugging Face Build Small hackathon (Thousand Token Wood). Everything runs locally — no cloud APIs.
How it works
your webcam ──(only on change, never stored)──▶ ① SEE: a vision model captions the frame
② JUDGE: a small LLM voices Mochi & Loaf,
situation-aware (you appeared / left /
are eating / on your phone / in the dark)
and grudge-aware (running gags)
both run in-process via llama.cpp — no API- Drag any image onto the page and the cats will judge it (your lunch, your dog, a meme).
- Dim your lights → their pupils dilate like real cats. A fresh verdict → Mochi pounces.
- Share their verdict copies a postable caption.
- Open the 🧠 HIVEMIND trace to watch the whole pipeline think.
Models (all local, via llama.cpp — no cloud)
Total well under the 32B cap. Swap models with env vars (VLM_REPO / COMPANION_MODEL_REPO). A deterministic stub keeps the cats funny even before the models warm up.
Badges: Off the Grid (no cloud APIs) · Llama Champion (llama.cpp runtime) · Off-Brand (custom three.js-free canvas/SVG frontend) · OpenBMB (MiniCPM-V + MiniCPM3).
Run locally
pip install -r requirements.txt
./run.sh # all-OpenBMB: MiniCPM-V eyes + MiniCPM3-4B brain
FAST=1 ./run.sh # moondream2 eyes + Llama-3.2-3B brain (snappier)
# open http://localhost:7860On Apple Silicon, install the Metal build of llama-cpp-python for fast local inference (--extra-index-url https://abetlen.github.io/llama-cpp-python/whl/metal).
Credits
Built with Llama (Meta) and MiniCPM / MiniCPM-V (OpenBMB), run via llama.cpp / llama-cpp-python. Webcam frames are described on-device and never stored or uploaded.