Private work · proprietary

POC_RETAIL_ANALYTICS

Private repository. Source available on request.

  • Licence not published
  • Status hibernated · 2026-10-11
  • Runs on Radxa Cubie A7Z
A row of detections, one of them tracked.
Illustration: a grid of pale dots on a dark plane, five of them lit in lime green along one path, with cyan lines radiating from a central point.

Stack

Neural Retail Mesh: an edge analytics engine for physical retail. Cameras and sensors on each node detect and re-identify people locally, and the nodes share vector embeddings rather than images, so a visitor can be followed across the floor plan without leaving the store.

A central siphon discovers the nodes over mDNS, fuses their streams and handles cross-camera deduplication. The edge runtime is a single Python process of about 75,000 lines with the dashboard embedded in it; the siphon is a separate service.

  • Python
  • YOLOv8
  • OpenCV
  • ONNX Runtime
  • Radxa · ARM64
  • PostgreSQL
  • mTLS
  • Zeroconf

The repository is private and the licence is not published. What is published here is what its own README documents: the architecture, the stack, and the date the work stopped.

State, measured from the repository
Last commit 2026-10-11 · 1d0f4c4 on main
History 288 commits
Status Hibernated. Its STATE.md records the last session as 2026-10-11 and the branch as the original baseline monolithic version.
Licence Not published. The README asserts copyright; there is no licence file.