Multi-Agent Tornado Crisis Coordinator

A multi-agent AI system for tornado crisis management and relief coordination. Four LangGraph-orchestrated agents combine CNN radar detection, LoRA-tuned tweet classification, and retrieval-augmented generation over FEMA guidelines to deliver real-time, explainable decision support for disaster response.

Coordinating relief in real time

When resources are scarce and time matters, disaster response is a coordination problem: the signal is scattered across radar feeds, social media, and hundreds of pages of official guidance. This system fuses all three into a single structured briefing an emergency operator can act on, and every recommendation is traceable to its source.

Architecture

Four specialized agents run under LangGraph orchestration:

Radar agent scores tornado likelihood from NEXRAD imagery with a custom lightweight CNN (focal loss, 15-channel engineered radar features) trained on TorNet volumes from 2013–2022, where tornadoes are only ~6–7% of samples.
Tweet agent triages crisis-related social media with DistilBERT + LoRA adapters, doing binary on-topic filtering and multi-class categorization; training labels come from Snorkel weak supervision over tweets from major U.S. tornado events.
Resource coordination agent retrieves relevant passages from ~11 FEMA manuals via Sentence-Transformer embeddings and a FAISS index, so recommendations are grounded in official guidance rather than model priors.
Summarizer agent synthesizes everything into a structured JSON briefing for operators.

Design choices

The system is tuned for the life-critical setting: recall over raw accuracy for the detection path (a missed tornado costs far more than a false alarm), explainability through RAG grounding (every recommendation cites FEMA guidance), and structured machine-readable output so downstream tooling can consume briefings directly.