Systems software in pure Rust — model quantization, real-time computer vision, and edge AI.

Core Epoch LLC is an independent software and IP company. We build software across a range of industries, but our main work centers on quantization — running computer vision models smaller and faster than full precision on CPUs, with no GPU or cloud service required.

Our production source is proprietary; however, we offer some of our models for free here and on Hugging Face. Technical write-ups, benchmark packages, and source access are available under NDA: contact@coreepoch.dev.

Kenosis is a pure-Rust quantizer for ONNX computer vision models. Every model it produces runs on both ONNX Runtime and OpenVINO from the same file.

• ImageNet classifiers: 33–58% less inference time on stock ONNX Runtime, 84–89% less on OpenVINO, versus full precision
• COCO detection: 93–94% of FP32 mAP retained at 38–50% less inference time (PP-YOLOE+)
• RF-DETR transformer: 48.4 mAP against 28.9 for ONNX Runtime's own static quantizer at its default calibration
• Research published with DOI: 10.5281/zenodo.20657989

Measured on an Intel i5-13420H. Four quantized models are published on Hugging Face, including an ImageNet-1K classifier that scores 81.53% top-1 in a 7.2 MB file.

Cryphex runs continuous object detection and tracking across multiple camera feeds, using Kenosis-quantized INT8 models, with no GPU required.

• Twelve 720p cameras with the detector gated to 5 fps each — the cadence NVR stacks recommend, run on every camera rather than behind a motion trigger: 29.8 fps of video per camera, 59 detections per second, 6.4 CPU cores, 722–730 MB of engine memory
• With the detector gate removed — every arriving frame detected — the same 320 px tier holds two cameras at full rate on 2.5 CPU cores, and 90–97 detections per second aggregate across four to six streams
• Six cameras on the 640 px accuracy tier, on OpenVINO: 24.2 fps aggregate at ~7 CPU cores; the full-precision model manages 13.1 on the same budget
• The 640 file scores the same 40.6 COCO mAP on ONNX Runtime and OpenVINO

Measured headless on an Intel i5-13420H; the twelve-camera row is the default configuration, the mean of three repeated steady-state passes. A zero-loss configuration is also available — in repeated measured passes, nothing taken in was lost between intake and output — at about 27% more CPU. Deployments are calibrated on your own hardware: frames are sampled on site, the model is re-quantized on the box, and nothing is uploaded.

Urim Lens turns any 2D video into a depth-driven 3D scene as it plays, in real time on the GPU.

coreepoch.dev · cryphex.dev · contact@coreepoch.dev · Charles Town, WV