Zorky: computer vision for finding orders in retail
A computer vision SDK and admin panel for the store associate's mobile app. The associate points the camera at a shelf and Zorky finds and highlights the right order or shoebox within seconds. Even offline.
Challenge
Pick-up point and store staff search for orders by eye, going through dozens of boxes and labels. It is slow, costly and leads to handover errors, and during sales search time grows faster than the load. Cloud recognition is not an option: connectivity in warehouses and back rooms is unreliable, and camera frames must not leave the perimeter.
Solution
We built Zorky — an on-device model that recognizes numeric codes on labels in the video stream and highlights matches right in the frame. The module plugs into the existing associate app via a single Kotlin Multiplatform SDK for Android and iOS, and a multi-tenant admin panel manages users, devices, ML models and quality metrics.
Orders by last 4 digits, shoes by 3 article digits, whole sector with a “Found X of Y” counter
The model runs on the device; frames never leave the phone — only anonymized metrics are sent
One public API for Android 24+ and iOS 15+, CameraX and CVPixelBuffer adapters, private Maven
Users and roles, device fleet, model versions with A/B and rollback, analytics and reports
Results
- Recognition in ≤2 seconds regardless of how many orders are on the shelf
- Target precision and recall of 95%+, IoU of 0.85+ on the required device matrix
- Runs on budget smartphones, including Huawei devices without Google Mobile Services
- Enterprise-ready: multi-tenancy, SSO, audit log, SBOM, SAP integration, GDPR
- Product website: zorky.softverno.ru