E-Buyur Market: Fruit and Vegetable Quality Detection Platform Using On-Device Hybrid Artificial Intelligence Method to Reduce Food Loss in MSMEs

Abstract

Indonesia produces large volumes of fruits and vegetables yet faces high food loss and waste (FLW), with studies estimating 23-48 million tons of food waste annually. Observations in Grobogan Regency show traders discarding around 30-40% of daily stock, often produce that is edible but visually imperfect (grade B/C). This study designs and evaluates E-Buyur Market, a digital marketplace embedding an on-device hybrid AI pipeline to automatically signal fruit-vegetable quality. The system combines a YOLO-TFLite commodity detector with a multitask MobileNetV3-Small TFLite grader to predict type, edibility, and a quality score derived from freshness probability. The model was trained on approximately 115,000 images compiled from publicly available Kaggle fruit-vegetable repositories, processed using data augmentation and post-training quantization for TensorFlow Lite deployment; a separate small set of real seller-captured photos was used to qualitatively test the deployed model under real usage conditions. On a held-out test set, the model achieves a type-classification accuracy of 0.91, freshness ROC-AUC of 0.94, PR-AUC of 0.92, and inference latency of tens of milliseconds on mid-range phones. Based on the system's design and signaling-theory rationale, on-device hybrid AI is expected to improve sell-through, reduce shrinkage, and shorten time-to-list for MSME sellers; empirical validation of these business-level outcomes through a live field trial remains future work.