Generative AI Based Mobile Application for Vehicle Breakdown and Maintenance Assistance

Vehicle maintenance presents significant challenges, particularly for those lacking mechanical expertise or access to professional services. While Google and ChatGPT offer basic assistance, they lack specialized automotive knowledge and real-time accuracy due to training data limitations. Existing applications like YourMechanic Mobile Car Repair primarily focus on service scheduling rather than repair guidance, with limited availability in regions like Sri Lanka. This research introduces MechAssist, a mobile application powered by fine-tuned Large Language Models (LLMs) for real-time vehicle diagnostics and maintenance guidance to address this gap. To identify the optimal machine learning model, five pre-trained models - FLAN-T5, NVIDIA-mistral, Facebook OPT-125M, GEMMA-2B, and BART - were evaluated using performance metrics after training with the dataset. The selected model was uploaded to the Hugging Face repository, cloned to a hosted server for ChatBot development, and subsequently integrated into the mobile application. User acceptance testing and customer feedback were then collected.The model selection was based on three primary criteria, which are the model size post-training, output response time, and output quality. BART demonstrated superior performance with an average response time under 2 seconds and a model size of 1.5GB compared to alternatives. Customer satisfaction with the application was notably high. The research concludes that BART is most suitable for this use case, and applications like MechAssist are valuable for vehicle owners with limited maintenance knowledge. Future development plans include expanding the dataset to 50,000 records, incorporating computer vision capabilities, implementing multi-language support, and enabling offline functionality through advanced TensorFlow Lite optimization.