ReGenLab

Vigil is a virtual safety and security officer for factories, built on Meta’s models and running on the CCTV cameras a factory already has. Most factories record their shop floor, gates and storage, but nobody watches it. A worker without a helmet near machinery, a blocked fire exit or a gate left open overnight is found only after an incident or in a manual footage review. Meta’s models do the thinking at every step Muse Spark (Meta Model API) is the brain. It looks at each frame and a crop of every person and judges every safety rule for every person, with a confidence and an explanation. Muse Spark builds the scene log. From the same look it reports, per person, which PPE is worn, and whether each gate or fire exit is open, closed or blocked, in each zone drawn on the camera. Muse Spark understands the rules. A safety officer types a rule in plain English (“alert if anyone is in the press area without a helmet”). Muse Spark compiles it into a schedule, a zone and a minimum duration, which the code then enforces exactly. The officer confirms the reading before it goes live. Muse Spark double-checks itself. A borderline finding is re-judged on six frames from a short clip, with only the suspect boxed, before any alert is sent. That is how Vigil keeps false alarms low. Meta’s SAM 2.1 can cut each person out of the frame so Muse Spark judges one clearly lit person at a time (optional stage). Meta’s WhatsApp Cloud API delivers the alert: the video clip with the person boxed and the zone shaded, delivery and read receipts, and a daily safety report. A cheap local layer (YOLO11-pose + ByteTrack, plus motion gating) only decides when and where to look: it finds and tracks people with persistent IDs so Muse Spark is called for what matters, not for every frame. It connects to the existing DVR/NVR over RTSP, so no new hardware is needed. Measured, not assumed: staff mark each alert real or false and report events Vigil missed. A Metrics page shows precision, recall, false alerts per day, alert latency and Muse Spark token cost for each camera. Privacy: no face recognition. Only alert events and short clips need to be stored; full-footage recording can be switched off, and old clips are deleted automatically after N days. Built with FastAPI, React, go2rtc and Backblaze B2 on a Vultr GPU server. Working end to end on recorded CCTV-style footage; accuracy figures will come from the reviews the Metrics page collects.