The flow is simple:

Camera (RTSP) ──→ Tensor Dynamics Watch ──→ AI model ──→ Live log
                                       (count / anomaly / threat)

• You provide: an RTSP URL per camera.
• The app samples video at the configured interval. The VLM default is one frame every five seconds, separate from object-detection refreshes.
• The AI returns: person count, scene description, and whether to alert (the tile turns red).

Capacity

The desktop supports a grid of up to 64 feeds; the API source kit defaults to 8. Actual capacity depends on the model, hardware, video, and site testing. Start with representative cameras and validate the setup before scaling.