Plant safety is a supervisor console for a warehouse floor. It turns camera clips into a queue of incidents: someone off the marked walkway, a forklift carrying three or more blocks, an open machine panel, or a worker at a machine without a high-vis vest. Each case shows the clip, why it was flagged, a coaching action, and who is assigned to investigate it.
A clip is ingested into the team video archive. The app searches that archive and builds an overview: how many incidents were flagged in the selected period, how many still need an owner, how many are under investigation, and how they split across severe, high, medium, and low.
VAST is the system of record. Its video index stores the segments and finds moments by meaning. Playback streams those segments. When a supervisor assigns a case, the owner and status are saved in a VAST database table, so the assignment remains after a restart.
Cosmos watches each clip and writes the caption on the case. The ingest prompt asks for literal facts: on or off the green path, the load count, whether a panel cover is open, and whether the person at the machine is wearing a vest. Safety rules read that caption.
Weights & Biases classifies the same caption when a rule is unsure. The app calls the W&B inference API with Llama 3.1 8B and asks for one safety class. A confident rule still wins, and the case says whether W&B agreed.
SpaceXAI (Cursor)
CoreWeave (Weights & Biases)
NVIDIA