LLM-Powered HFACS Analysis of ASRS Narratives via MCP Bridge for Enhanced Aviation Safety Debriefs

Learning effectively from past incidents is crucial for advancing aviation safety, but manually analyzing extensive narrative datasets like the Aviation Safety Reporting System (ASRS) to understand causality is resource-intensive and can be subjective. While Large Language Models (LLMs) offer powerful capabilities for text analysis, their performance can degrade when processing large datasets within limited context windows that could potentially lead to decreased accuracy and increased hallucination. This paper introduces a novel framework to address these challenges by automating the extraction of human factors insights from Oklahoma ASRS records. We employ the API version of OpenAI's o3 reasoning LLM, guided by few-shot prompting, to classify ASRS incident narratives into the Human Factors Analysis and Classification System (HFACS), each record at a time. The core of our solution is a specialized Model Context Protocol (MCP) server designed to intelligently serve this data. We enable targeted data retrieval by using MCP, which allows the LLM to focus on relevant information from the ASRS corpus without overwhelming its context capacity during complex queries. This server is integrated into a broader ecosystem via the MCP Bridge, an LLM-agnostic RESTful proxy, which further facilitates interaction with diverse clients. This agent demonstrates the system's capability to perform complex, multi-tool, zero-shot queries to retrieve, compare, and report on incidents based on specific HFACS criteria and keywords. Furthermore, to address the computational costs of large models, we introduce a specialized 8B-parameter agentic model trained using Group Relative Policy Optimization (GRPO). This fine-tuned model demonstrates increased performance in strict argument generation (71.3%) compared to a 70B-parameter general-purpose model (54.3%) that proves smaller, reinforcement-learning-optimized models can outperform larger counterparts in specialized tool-use tasks. Our approach highlights how combining advanced reasoning LLMs, specialized agents, and MCPs’ targeted data access can significantly enhance the objectivity and efficiency of safety investigations.