During our fifth week, we reached a major milestone by finalizing the State of the Art section of our report. This comprehensive review of existing literature and solutions solidifies our theoretical foundation and justifies the technical approaches we are taking to tackle insurance fraud detection.
In the Computer Vision module, we transitioned from testing baseline pre-trained models to actual fine-tuning. By adjusting the model architectures and weights to our specific, pre-processed dataset, we aim to significantly improve the accuracy and reliability of damage and anomaly detection in vehicle images.
Meanwhile, in the Natural Language track, we focused on scaling our custom dataset. To achieve this, we developed a dedicated API integrated with a Generative AI model. By engineering a sophisticated and highly specific prompt, we are now able to automatically generate diverse, high-quality synthetic accident descriptions, massively enriching our training data and preparing our NLP models for the next stages.