CrisisFlow

Inspiration

Disaster response fails not from lack of data — but lack of coordinated data. Hospital ERPs, supplier portals, weather feeds, and government incident systems all operate in silos. When a wildfire hits, emergency coordinators spend 72 minutes manually stitching these together. 34% of the time, supplies arrive after the patient surge already peaked. We built CrisisFlow to close that gap.

How we built it

Challenges we faced

What we learned

What it does

CrisisFlow is an AI-powered emergency medical supply coordination agent. From a single incident signal, it predicts patient surge, computes real-time resource gaps, queries a pre-vetted supplier network, and generates a ranked dispatch plan, getting ready for human approval in under 5 minutes.

How we built it

We used Google Gemini as the core reasoning agent, with Fivetran → BigQuery as the data pipeline connecting hospital inventory, supplier contracts, weather feeds, and government incident feeds. The dispatch pipeline runs as a 6-step autonomous agent: Detect → Predict → Analyze → Source → Plan → Dispatch. Every high-risk action requires explicit human approval before execution.

Challenges we ran into

The hardest problem was supplier query reliability, real emergency networks have inconsistent APIs, partial data, and fallback-only integrations. We built a tiered integration model (API → portal → email/SMS) to handle degraded conditions gracefully. Balancing agent autonomy with human oversight in life-critical decisions was also a core design constraint we had to get right.

Accomplishments that we're proud of

Turning a 72-minute manual process into a 5-minute approved dispatch plan — end to end, fully automated except for the human approval gate we intentionally kept. Building a tiered supplier integration that degrades gracefully: API-first, portal fallback, SMS last resort, so the system stays functional even when infrastructure is stressed, which is exactly when you need it most. And shipping a working multi-agent pipeline under hackathon conditions, where the hardest constraint wasn't the tech. It was designing for life-critical decisions where a wrong recommendation has real consequences.

What we learned

Multi-agent coordination in high-stakes environments requires explainability at every step, not just a recommendation, but why that supplier, why that route, and what happens if it's delayed. We learned to design for the approver, not just the algorithm.