AI-Augmented Catastrophe Modeling with an Intelligent Orchestration Workflow
Conference
Patel, JA, Wang, T, Hamid, S et al. (2026). AI-Augmented Catastrophe Modeling with an Intelligent Orchestration Workflow
. 292-297. 10.1109/IRI69576.2026.00062
Patel, JA, Wang, T, Hamid, S et al. (2026). AI-Augmented Catastrophe Modeling with an Intelligent Orchestration Workflow
. 292-297. 10.1109/IRI69576.2026.00062
The Florida Public Hurricane Loss Model (FPHLM) estimates hurricane-induced insurance losses from exposure datasets submitted by multiple vendors. Although these datasets are expected to follow a canonical specification, in practice, they often contain heterogeneous schemas, malformed fields, and recurring data-quality defects that require extensive manual intervention. This paper presents an AI-augmented orchestration workflow for FPHLM that integrates agentic AI with explicit human oversight to improve preprocessing, monitoring, and verification while preserving operational safety and auditability. The proposed framework combines an LLM-guided preprocessing layer, a retrieval-augmented generation (RAG)-based autoverifier, and an anomaly-detection module for identifying unusual policy-file behavior and pipeline outcomes. Rather than treating large language models as a mere coding aid, the system uses them as operational components within a constrained multiagent protocol to support structured data cleaning and grounded verification. The contribution of this paper is the enhancement of automating the cleaning and verification workflow in the CatOP framework, a dynamic alerting module, and outlier detection using the DBSCAN for detecting abnormal processing patterns that may indicate data or workflow inconsistencies.
