Hours of raw footage from a conflict zone. Five seconds that matter. Real-time explosion alerts for the news desk, and the same agent watching streets and warehouse floors.
222 hand-checked 5 s clips from five films: B-17 raids, the RAF attack on the Tirpitz, RAF night bombing, USAAF "Special Delivery" (1946) and "Disney Bomb" (1945).
10% precision. Parachutes and supply drops read as bombs; a stricter prompt only reached 12%. We show it, not hide it.


Wikimedia Commons, public domain. Sent to Cosmos directly for evaluation; not ingested into the VSS archive.
Truth: human frame check of every segment. Wikimedia Commons, public domain; sent to Cosmos only, not ingested into VSS.
Indexed on VAST: Toronto dashcam, a New York cyclist's helmet camera, NYC and San Francisco street cameras, a highway, a neighborhood camera.
Forklifts, AGVs and humanoid robots sharing space with workers.
Where did a person come dangerously close to a moving vehicle or machine?
A ranked, human-verified incident feed with bounding boxes, plain-English alerts and written reports.
Ingests and segments video into 5 s clips; YOLO boxes, captions and embeddings in VastDB.
Watches each clip (on CoreWeave GPUs) with a street or warehouse safety prompt: risk 0–5, event, machine, people nearby.
Consistency rule (no human near the machine means no alert); human confirm/dismiss with notes.
gpt-oss-120b writes reports and compiles watch rules; Weave traces every call; eval runs measure accuracy.
Web app deployed on the team Kubernetes namespace behind the /app Ingress. Open it from workshop.thecosmoslabs.com → App.
| Check (human frame review = ground truth) | Result |
|---|---|
| Street clips: Cosmos risk vs. reviewer | 8 / 10 agree |
| Warehouse clips (after the zero-humans consistency rule) | 6 / 6 agree |
| Archival war film (5 films, 222 clips): explosions | 85% precision · 91% recall |
| Live trigger, 3 unseen films (36 segments): explosions | 3 / 3 blasts alerted · 0 false alarms |
| Archival war film: bomb release | 10% precision |
Every Cosmos Reason and LLM call is traced in Weave (project jayakumaronline/vision-zero-agent). Results are logged as the W&B run "cosmos-reason-frame-checked-eval". Reviewed street and warehouse clips are spot checks, not a random sample.
Video storage, semantic search, Q&A over the archive, YOLO detections, streaming.
Cosmos Reason watches every clip and scores the risk.
Serverless inference for reports and watch rules; Weave tracing; evaluation runs.
Built, deployed and tested with Cursor agents, subagents and skills.
Any camera, any archive. Every finding checked on the video, and every number measured.