Frappua

AI Vision Monitor: Safety Mode. Write a workplace safety rule in plain words, and an ordinary camera checks it. A safety manager types "Everyone on the warehouse floor must wear a hard hat and a high-visibility vest." Muse Spark compiles the sentence into a strict JSON rule and shows what it understood: objects, zone, condition, hold time, severity. If a camera cannot check the rule, Spark refuses with a reason and a wording that works. SAM 3.1 finds the objects in each frame, with track ids in video mode. Our zone engine decides when to alert, so the same detections always give the same alerts. Spark explains each alert and writes the shift report, and our code rejects any text with a number that is not in the event data. Rule Scout, a Spark agent with tool calls, proposes rules for a camera, runs them through the engine and keeps the ones whose alerts hold up. On site, offline: the rule compiler runs on Muse Glimmer 30B, Meta's open-weight model, on a Jetson AGX Orin 64 GB. It compiled all 10 reference rules into our strict schema, 53 s per rule on average, 18 GB of memory. The detector SAM 3 runs on the same board: 873 ms per frame with 3 prompts, 2 cameras at one frame every 2 s. Video never leaves the building. How to test Live demo, no sign-in and no key: https://safety-mode.vercel.app 1. Your own clip: open "Your clip", choose a video, draw a zone on the first frame with the mouse and name it, then write a rule that uses the name, e.g. "Everyone at the desk must wear glasses." The file stays in your browser; SAM 3.1 gets one frame per second of the first 30 s. 2. Webcam at your desk: open "Webcam", allow the camera, press a "Try at your desk" rule (glasses, headphones, no cup in the keyboard zone) and break it. The alert comes on the first frame that shows it. 3. Any rule in plain words on the sample clips, e.g. "Nobody may stand on the shop floor without gloves." SAM 3.1 runs live for objects we have no cached detections for. The four example rules on the sample clips, their explanations and recorded Rule Scout runs come from our cache. New rules, new objects, the webcam, your clip, the report and a live Rule Scout run call the Meta API live, within hourly limits per address. AI Vision Monitor is Frappua's existing camera alert product; its code is not in this repo. Built during the hackathon (26 Oct to 4 Nov): everything in safety_mode/, tests/, web/, api/, demo/ and jetson-bench/glimmer.md.