Primate Vision ingests live video over WebRTC — the same protocol that powers video calls — and returns analysis results on a data channel while the stream is still running. WebRTC gives us sub-second ingestion latency, adaptive bitrate, and broad device support without requiring a separate streaming infrastructure layer. You point a camera (or stream a file at live speed) and ask a question; results appear as the video plays.
On the backend, each active stream is assigned to a per-GPU worker that runs Darwin's JEPA encoder continuously. The key trade-off is time-to-first-frame (TTFF) versus sustained throughput: a cold stream worker takes longer to produce the first result than a warm one. We publish p50 and p95 TTFF numbers on theperformance page. Sustained streaming analysis runs at 11.8 fps on production hardware, with 45ms p50 per-frame inference. "Real-time" means the analysis keeps up with the video — not that it completes before the frame arrives.
See also: Streaming guide in docs.
The full engineering story — WebRTC internals, queue architecture, and edge deployment — coming during our August launch.