Inspiration
Florida faces some of the most powerful hurricanes in the U.S, leaving civilians and businesses vulnerable. In moments of crisis, people need clarity, not scattered information. Our inspiration was to mitigate the chaos of finding trustworthy information and resources tailored to their specific location by helping people focus on safety instead of scrambling for answers.
What It Does
Our platform displays historical hurricane paths alongside predicted tracks and then launches a multi-agent pipeline tailored to the user’s location. Once a ZIP code or coordinates are entered, the system kicks off:
Forecast & Alerts: Real-time hazard snapshots from NHC + NWS.
Checklists: Personalized supply lists based on FEMA/Red Cross guidelines.
Resource Finder: Verified sandbag, pantry, and shelter locations based on County prompt.
Routing: Safe, unblocked routes to supply stores using live closure feeds.
Home Prep Advisor: A multimodal Gemini-powered agent that analyzes user-uploaded home photos, detects vulnerabilities, and returns prioritized actions with confidence scores.
News Curator: Latest local updates and official advisories.
Business Tools: Cost modeling, continuity planning, and executive summaries for organizations.
How We Built It
We designed a multi-agent architecture with clear division of labor:
Orchestrator: The entry point that fans out requests and aggregates results.
Data Agents: Fetch storm cones, alerts, and forecast snapshots.
Supply Planner: Uses heuristics (3L water per person per day, 2000 calories/day, and more) to build checklists.
Routing & Resource Agents: Map supply runs and official distribution points, factoring in closures.
Vision Agent: Uses Gemini for multimodal risk analysis with JSON Schema enforcement.
Business Agents: Estimate costs, optimize HQ sites, and create one-pager briefs with charts.
We leveraged:
APIs: NWS, NHC GIS layers, FEMA NFHL, FL511.
LLMs: Google Gemini (multimodal), local translation for paused NWS multilingual alerts.
Math/Heuristics: Haversine distance for routing, FEMA/Red Cross heuristics for supply planning.
Challenges We Ran Into
Data Fragmentation: Official resources are spread across dozens of county/state pages with no central API. We built crawlers + parsers to standardize this.
Multilingual Alerts: NWS paused translations, so we integrated local translation inside the chatbot.
Demo Reliability: Balancing cutting-edge agents (e.g., Vision risk assessment) with fallback hardcoded data for live demo stability.
Agent Coordination: Designing schemas so agents could pass structured outputs (JSON) reliably without conversational noise.
Accomplishments We’re Proud Of
Built an end-to-end pipeline where multiple autonomous agents cooperate in real time.
Created a Vision-powered risk advisor that can look at your home and give concrete prep actions.
Delivered both civilian and business workflows within one system.
Demonstrated Agent2Agent (A2A) and local orchestration working seamlessly in parallel.
What We Learned
The importance of schema enforcement for LLM outputs—agents must “speak JSON,” not prose.
How parallelism can speed up disaster response (supply planning, routing, and sandbag search run simultaneously).
Balancing reliability vs innovation in a hackathon demo—fallbacks are essential.
The complexity of trust in data sources: users need to know alerts come from NWS/NHC, not a random scraper.
What’s Next for Cvclone
Expand beyond Florida to cover all hurricane-prone states.
Integrate real-time satellite feeds and ensemble weather models for sharper forecasts.
Deploy a mobile app with offline caching for when power/internet fail.
Partner with counties and aid orgs to ensure official sandbag/shelter info flows directly into the platform.
Extend the business side to support insurance firms and logistics networks with automated continuity playbooks.
Log in or sign up for Devpost to join the conversation.