Recall Recon: Predicting Automotive Safety Recalls with AI

Recall Recon is an AI-driven system developed to forecast automotive safety recalls before they happen. The project addresses a major gap in the vehicle safety landscape, where recalls are typically issued only after failures have occurred. Despite advances in vehicle diagnostics and autonomous systems, the current recall process remains largely reactive. Recall Recon proposes a proactive model using historical safety data and machine learning to identify patterns that precede large-scale recalls.

At the core of Recall Recon is a Random Forest regression model trained on decades of National Highway Traffic Safety Administration (NHTSA) data. This includes over 70,000 recall records across multiple components like airbags, brakes, and electrical systems. To improve accuracy and interpretability, the team engineered several domain-specific features such as recall completion rate, seasonal indicators, and manufacturer risk scores. These enhancements allowed the model to forecast recall volumes with a mean absolute error of about 4.7 million vehicles.

To complement the forecasting model, the team built a Retrieval-Augmented Generation (RAG) chatbot for natural language interaction. This tool allows users to ask detailed recall-related questions in plain English. The system retrieves relevant documents using vector search (FAISS) and generates responses with OpenAI’s GPT-3.5 Turbo. It provides fast, explainable answers to diverse queries, making complex recall data more accessible to analysts, manufacturers, and regulatory bodies.

Visualization tools were also developed to support decision-making. Power BI dashboards offer strategic views of recall trends, enabling OEMs to monitor high-risk categories, allocate resources, and plan service campaigns. A Streamlit interface allows users to explore interactive charts and receive real-time chatbot responses, making the system useful for both executive overviews and technical deep dives.

The system was evaluated for both predictive performance and conversational reliability. The RAG chatbot achieved 91.3% Precision@5 in internal testing and maintained 90% factual accuracy across varied queries. The forecast model revealed consistent recall volumes in the 40 to 50 million range annually, highlighting the persistent nature of systemic safety defects, even in modern vehicles.

Recall Recon demonstrates how machine learning and language models can be combined for real-world impact in safety-critical domains. Future plans include VIN-level recall prediction, voice assistant integration, and global data fusion. By making recall intelligence predictive, interpretable, and actionable, the project sets a strong precedent for using AI to enhance automotive safety.