AI-Assisted Product Identification and Decision Support for Industrial Warranty Claim Management

Sammanfattning

Warranty claim processing at ELAJO Elektriska AB depends entirely on manual product identificationand paper-based documentation – a workflow that is both time-consuming and prone to error. This thesispresents an AI-assisted system that automates product identification from field photographs and integratesthe result into a digital warranty claim workflow.The system first attempts barcode decoding. If no readable barcode is present, DINOv2-Large encodesthe photograph into an embedding vector and searches a FAISS index of 12,008 pre-computed catalogueembeddings – expanded from 1,501 products using photorealistic augmentation– to retrieve the 30 mostsimilar candidates. CLIP ViT-L/14 re-ranks this shortlist and the top-5 results are presented to the technician for confirmation. An order-filtered search mode further improves accuracy by restricting retrievalto only the products belonging to the technician’s specific order, achieving identification accuracy above95%in this setting.Five pre-trained models were evaluated under identical zero-shot conditions on 68 real-world field photographs, each tested with and without augmentation: ResNet-50, EfficientNet-B4, CLIP ViT-L/14,DINOv2-Large, and a DINOv2+CLIP ensemble. The DINOv2+CLIP ensemble achieved the best Recall@1 of 50% and Recall@5 of 75% . Augmentation improved DINOv2 accuracy by 7.37 percentagepoints without any model retraining.The system is delivered as a full-stack web application with a React frontend and FastAPI backend connected to a Microsoft SQL server database. It checks ALEM09 warranty eligibility automatically andgenerates a pre-filled digital claim form for technician review and submission. The source code is publicly available at https://github.com/AmeJob-AI/AI-Product-Identification.