OverWatch runs its detector, vision, and language models locally in your browser via WebGPU (with WebGL/WASM fallback). On first visit these models download from the Hugging Face Hub — this can take a moment on a cold cache. Once cached, subsequent loads are instant. Telemetry and connectors work independently while models load.
YOLO + COCO-SSD (TensorFlow.js) — Web Worker on WebGPU, falls back to WebGL/WASM.
LFM2.5-VL vision-language model (@huggingface/transformers) — captions camera frames on-device.
SmolLM2-360M-Instruct — powers the analyst chat & briefings. Loads on first prompt.