The system runs continuously on an edge server connected to CCTV cameras via RTSP streaming. Runs RF-DETR person detection, applies spatial exclusion zones (doorways, equipment), and computes a productivity score from worker presence and background-subtraction activity.
Violation of specified rule triggers a Telegram alert with an annotated frame after a configurable number of confirmation cycles. At the end of each day a vision-language model reviews the hourly snapshots and generates a written summary exported as a PDF report.
