A working slice of your pipeline — in your stack, running today.
Motion-triggered detection at fleet scale: the camera is an edge pre-processor, not a video firehose. We rebuilt your path — ONVIF → detection → interpretation → operator — in Python/FastAPI, with AWS mocked so it runs anywhere.
The whole path, one trace.
Seven beats: fleet at scale, onboard a camera with no deploy, motion live. Then the Protocol Trace — real per-stage spans and latencies on one trace_id worker→backend — operator ack/resolve, resilience (429 → DLQ → fallback; kill a camera, neighbours hold), and the quality gate.
The three workstreams.
Registry & provisioning
Onboard a camera by DB write — a worker spawns, streams live, no deploy. RTSP creds go to a secret store, injected at spawn.
DeepAlert adapter
POST /v1.0/analytics mapped to your detection contract, with client-minted job_id dedup. ROI, keyframe, and interpretation run unchanged.
Resilience
On a 429 we retry, then hold to DLQ and fail over through the circuit breaker. Per-camera health isolates faults — kill one camera and the neighbours keep streaming.
Does this match what you need built?
This was built in the days after the scoping call, in your stack, mapped one-to-one to the 250-camera + DeepAlert MVP you scoped. If the approach lands, we'd attach per-workstream effort estimates and start where the risk is — the DeepAlert adapter on your existing single-camera worker.