Large Language Model Pipelines for Crisis Intelligence: Automated Help Request Classifier

Abstract

This research examines the development and deployment of a Crisis Intelligence Pipeline, engineered to address the catastrophic information surge following Cyclonic Storm Ditwah in Sri Lanka [1]. The system integrates advanced prompt engineering methodologies, specifically few-shot learning, Chain-of-Thought [2], and Tree-of-Thoughts [3] to transform unstructured humanitarian data into actionable logistics strategies. Through rigorous stability testing under varying temperature gradients, the study identifies a "Safe Mode" for deterministic triage and a "Chaos Mode" that highlights the risks of model hallucination in high-stakes environments. Furthermore, the pipeline utilizes Pydantic-based schema validation [4] and token economics [5] to ensure scalability and cost-efficiency. Results indicate that structured reasoning architectures significantly enhance the precision of resource allocation and signal-to-noise discrimination, providing a vital technological framework.