Detection models are quantized with Kenosis and verified against COCO ground truth. The verification measures detection accuracy itself, not a proxy metric.
Every camera runs in its own isolated thread. At real-time tiers the pipeline drops zero frames, and telemetry counter-verifies it.
The pipeline is native Rust and memory-safe by construction, with no Python layer in the serving path.
The same build needs no GPU, and it picks up CUDA and DirectML acceleration when the host has one.
Prometheus and REST monitoring are built in, with per-stage latency histograms and overload alerting. Every performance claim on this page is reproducible from the daemon's own counters.
Measured — multi-stream headless daemon, 720p sources, 2026-07-21
Deterministic 720p30 streams · PP-YOLOE+ S INT8 · Intel i5-13420H · 45 s windows · every configured stream verified at full ingest rate
Measured — multi-camera scaling, deterministic clip matrix 2026-07-21
Four to six cameras · 720p30 fixture clips over RTSP · PP-YOLOE+ S · 2 threads per camera · Intel i5-13420H · 45 s steady-state windows · FP32 shown on its faster runtime