CONTRA COSTA COUNTY EMERGENCY OPERATIONS CENTER · CALIFORNIA
This is what a real EOC looks like. Multiple data streams. High-pressure decisions. No room for bad UI.
Agentic AI Orchestration for disaster response in zero-connectivity environments.
It's 2:14 AM.
Hurricane Mara makes landfall.
The grid goes down.
THE SCENARIO
She runs the Emergency Operations Center
for a coastal county of 340,000 people.
Tonight is the worst night of her career.
THE PROBLEM / LIVED EXPERIENCE
Her dashboard — the one that tracks every volunteer, every supply truck, every shelter
— goes grey.
It runs on AWS.
AWS needs the internet.
A radio crackles: "Fire at the museum — need water, over."
She writes it on a Post-it note.
There are 47 other Post-it notes already.
She doesn't know where Unit Bravo is.
She doesn't know how much water is left at Depot 3.
She doesn't know if Shelter 6 is over capacity.
She's commanding blind.
"In major U.S. disaster declarations,
cellular infrastructure typically fails
within the first 6 hours of impact."
— FEMA INFRASTRUCTURE RESILIENCE GUIDELINES
Traditional emergency software fails exactly when it's needed most.
What if the command center
never needed the cloud
to begin with?
A resilient command center that runs entirely on-device. No Wi-Fi. No cloud APIs. No single point of failure.
When the internet stops existing, DisasterHub keeps working.
CORE FEATURES
Four interlocking systems, all running locally. Each one designed for a commander under pressure.
A radio burst crackles through static:
"Fire at the museum — need water."
The Comms Bridge Agent parses fragmented field reports into structured incident data in under 2 seconds. No internet. No cloud API call. Just local AI, running on the device in front of the commander.
The AI scans the local database. Unit Bravo is idle. Unit Bravo has water. Unit Bravo is 4 minutes away.
A mission proposal is drafted automatically — incident matched to resource, route calculated, ETA confirmed. All without a single server ping.
DisasterHub never deploys a unit without authorization.
Every AI-generated mission proposal surfaces as a reviewable card. The commander reads the logic, checks the map, and decides. [APPROVE] or [REJECT] — always one deliberate action away.
In GovTech, this isn't a UX detail. It's the entire philosophy.
When Med Kit inventory drops below 25%, the card turns red. Automatically.
But DisasterHub doesn't just alert — it calculates burn rate across all active incidents and flags procurement needs before items run out.
At 3 AM, no one is watching. The system is.
FEATURE 05 / VOLUNTEER TRACKING
Commanders can instantly correlate a volunteer's card — name, role, coordinates, phone — with their spatial position on the live map.
No tab switching. No searching. The master-detail layout keeps both views in peripheral vision simultaneously.
UNDER THE HOOD
Three technical decisions that made offline AI possible.
Phi-3 Mini runs fully in-browser via WebGPU using WebLLM. Voyage AI handles semantic embeddings for intelligent incident matching.
Zero cloud calls. Zero latency. Works at 2 AM with no signal.
A single AI model dynamically swaps System Prompts to act as:
→ Comms Agent (signal parsing)
→ Allocation Agent (mission drafting)
→ Inventory Agent (burn rate)
Memory-efficient. Contextually specialized.
Local-first database syncs volunteer positions, incident reports, and supply levels on-device.
Reconciles automatically when connectivity returns.
LIVE DEMO
A full walkthrough of DisasterHub handling a live simulated earthquake scenario.
DESIGN PROCESS
We iterated on the core layout once. It mattered.
The first version put the map everywhere. It felt cinematic. But when I watched someone try to use it, the problem was immediate — there was nowhere to look first.
No hierarchy. No at-a-glance status. Just a beautiful map that told you nothing.
I asked: what mental model does a commander already have? They're not navigating — they're monitoring. The map is one signal among many, not the whole picture.
I moved to a dashboard-first layout: persistent sidebar, glanceable metric cards up top, map as a supporting panel. Every element earns its pixel.
OUTCOME
IT'S 3:47 AM
Commander Reyes
approves her 12th mission.
Unit Bravo arrives in 4 minutes. The fire is out.
Shelter 6 is flagged before a single call is made.
The grid is still dark.
The map is still lit.
Designing for high-stress moments forced ruthless clarity. The "3 AM test" became our rule: can an exhausted dispatcher decide instantly? In disaster response, cognitive load is a real operational metric.
Building an AI orchestration tool with AI taught me one thing: expectation is the product. The interface isn't just UI — it's trust in invisible logic.
Next is true synchronization: one offline-sync system connecting the EOC dashboard and the volunteer app. To scale, I'd sync a real-time disaster database to map disasters as they occur.