YUNetLLMClaimReport: An Enhanced Automobile Insurance Fraud Detection and Automated Claim Report Generation using Large Language Models

The detection of fraud in automobile insurance and processing of automobile claims are still a challenge since manual inspections are carried out and automated systems are fragmented. The state-of-the-art approaches (convolutional neural networks, statistical models, etc.) tend to focus on separate tasks (damage detection, fraud classification, etc.), without being integrated with report generation on an end-to-end basis, and without making sufficient use of unstructured information (e.g., images, claim descriptions). In this research, YUNet_LLMClaimReport, a new framework is proposed that combines YOLOv11, U-Net, and a fine-tuned GPT-3.5-turbo large language model to automatically generate claim reports based on the detections and segmentation. The model is trained on a custom dataset of 4,000 vehicle images within the Egyptian insurance market (70:20:10 training, validation and test split), using data augmentation, dropout (0.3) and L2 regularisation (l = 0.01) to ensure robustness. The evaluation measures are mean average precision (mAP50: 0.9424), intersection over union (IoU: 0.85), claim report accuracy (CRA: 0.92 +/- 0.008), BLEU (0.85 +/- 0.01), and ROUGE-L (0.82 +/- 0.01). The framework runs at 8.3 FPS on NVIDIA Jetson Nano and saves processing time by 60 percent. Weaknesses are single-run measures because of computational limitations and factual inconsistencies in reports (10%). The model can be generalized to the Kaggle Car Damage Detection dataset (mAP50: 0.90), providing a scalable solution to insurance processes. The code and sample data are open sourced.