GondoCheck: A Vision-Based Backend for Automated Shelf Auditing

The persistent out-of-stock problem in Brazilian retail erodes revenue and customer loyalty. We present GondoCheck, a cloud backend that receives a single photo of a gondola, detects the stocked items, matches them to a target stock-keeping-unit (SKU) list through deep metric learning and returns real-time shelf-share analytics. Using 96 real supermarket images and six dishwashing-liquid SKUs, the system achieves 85.8% global accuracy and 0.77 macro F1, with 100% precision on two shelves after fine-tuning a ResNet-152 embedding. The approach shows that combining off-the-shelf object detectors with metric-learning embeddings can deliver practical, scalable shelf auditing without planograms or store calibration.