Zhiyu Quan
University of Illinois Urbana-Champaign
Room: M3 3127
Claim Automation using Large Language Model
While LLMs have achieved strong performance on general-purpose language tasks, their deployment in regulated and data-sensitive domains remains limited. Using millions of historical warranty claims, we develop a locally deployed, governance-controlled LLM system that generates structured corrective-action recommendations, labor estimates, and cost predictions from unstructured claim narratives. Domain adaptation via Low-Rank Adaptation (LoRA), together with a modular intermediate-task architecture, aligns the system with real-world claim workflows and supports regulatory governance and model risk management. We evaluate the system using a multi-dimensional framework that combines semantic similarity metrics and human evaluation. The results show that domain-specific fine-tuning substantially outperforms prompt-based approaches, with approximately three-quarters of the evaluated cases achieving near-identical matches to ground-truth corrective actions. Overall, this work demonstrates how high-quality operational data, domain-adaptive learning, and governance-aware system design enable responsible and impactful AI deployment in insurance.