Unlocking the most capable computer vision that needs no accelerator
Request an Evaluation →Detection models are compressed by our own quantization system, built for this problem rather than adapted from general-purpose software. The bundled model suite arrives already tuned, so a site gets full-accuracy detection on day one, across very different deployments and sectors.
Zero crashes logged to date. The core is Rust and memory-safe by construction, every camera runs on its own lock-free pipeline, and the hardening is deliberate rather than something we do after a failure teaches us to.
Cryphex is tuned to run efficiently where there is no accelerator to fall back on. Where a site already has one, it will use CUDA or DirectML natively.
Prometheus and REST monitoring are built in, with per-stage latency and overload alerting. Most deployments find they stop reaching for it, because the performance simply holds.
Video is processed where it is recorded. Models are calibrated on the customer's own hardware from their own footage, and nothing is uploaded.
Measured — twelve cameras on one mid-range CPU
720p30 streams over RTSP · three repeated steady-state passes · 48,333 frames delivered, none of them lost between intake and output
Measured — the detail tier, where accuracy leads
720p30 streams over RTSP · steady-state windows · the full-precision comparison measured on its faster runtime