Scramble

SCRAMBLE is Google Maps for disasters, AI that dynamically evacuates millions in real time using satellites, GPUs, and offline first alerts, turning chaos into coordinated survival.

Comment

Inspiration

The sirens rang. 230 million people heard them. And then—nothing.

No directions. No clarity. No plan.

Smart cities spent billions on sensors, alerts, and dashboards. But when disaster actually strikes, alerts don’t save lives.

Movement does. Coordination does. Intelligence does.

SCRAMBLE was born from a terrifying realization: Cities warn people—but they don’t guide them.

Evacuation isn’t a notification problem. It’s a flow optimization problem. And until now, no one was solving it.

What it does

SCRAMBLE is AI-powered evacuation intelligence for real cities, real people, and real chaos.

Not a broadcast. Not a generic alert. A live, thinking system that coordinates how millions move to safety.

SCRAMBLE:

  • Detects disasters in real time using live satellite intelligence
  • Calculates optimal evacuation routes per city zone
  • Dynamically balances people across shelters to prevent gridlock
  • Sends personalized, multilingual instructions via app and SMS (yes—even on 2G)
  • Gives responders a single, real-time command view of the entire city

Think: Google Maps × Disaster Response × AI — at city scale.

How we built it

We built SCRAMBLE as a real-time evacuation intelligence system, not a simple alert app.

Live flood data from Sentinel-1 SAR satellites is processed via Google Earth Engine and analyzed using GPU-accelerated AI on AMD ROCm. Flooded roads are automatically removed from the city’s road graph built from OpenStreetMap data.

An AI routing engine then optimizes zone-to-shelter assignments, balancing population flow and shelter capacity to prevent gridlock. Multilingual evacuation instructions are generated using Llama 3 and delivered via a citizen app and offline-first SMS, ensuring access even on low connectivity.

The entire pipeline—from satellite detection to citizen instruction—runs in under three minutes.

Technical Flow (End-to-End)

[ Sentinel-1 SAR Satellite ]
            ↓
[ Live Flood Detection AI ]
(PyTorch + :contentReference[oaicite:0]{index=0})
            ↓
[ Hazard Map Generator ]
(Flooded roads auto-blocked)
            ↓
[ AI Routing Optimizer ]
(NetworkX + A*)
            ↓
[ Zone → Shelter Assignment ]
(Capacity-balanced)
            ↓
[ Alert Intelligence Engine ]
(:contentReference[oaicite:1]{index=1} + vLLM, 5 languages)
            ↓
[ Delivery Layer ]
App + SMS (Offline-first)
            ↓
[ Responder Command Dashboard ]
(Live movement & bottlenecks)

From satellite detection to citizen SMS: under 3 minutes.

Because in disasters, minutes decide lives.

Architecture Overview

  • Satellite Intelligence: Sentinel-1 SAR via Google Earth Engine
  • AI Core: Flood detection + routing optimization on AMD ROCm GPUs
  • LLMs: Llama 3 generating actionable, multilingual evacuation instructions
  • Backend: FastAPI + PostGIS for city-scale geospatial intelligence
  • Frontend: React + Leaflet for citizens and responders
  • Offline Access: SMS fallback via Twilio / MSG91

Fully open-source. India-deployable.

Zero vendor lock-in.

Challenges we ran into

  • Turning raw satellite flood data into instant, actionable road closures
  • Routing millions of people simultaneously without collapsing shelters
  • Designing for low connectivity, feature phones, and digital exclusion
  • Making AI decisions transparent and trusted during emergencies
  • Achieving real-time GPU inference under extreme latency constraints

Every challenge forced us to confront one truth: Evacuation isn’t an app problem.

It’s a systems problem.

Accomplishments that we’re proud of

  • A complete evacuation intelligence pipeline running in under 3 minutes
  • AI-balanced routing across entire cities—not just individuals
  • A truly offline-first design that works when networks fail
  • Multilingual evacuation intelligence—not just translation, but clarity
  • Built entirely on open-source + AMD ROCm — no proprietary lock-in

But most importantly: SCRAMBLE works for everyone. Not just the connected. Not just the privileged.

Everyone.

What we learned

  • Evacuation is about movement intelligence, not louder alerts
  • AI must be infrastructure-aware, not cloud-dependent
  • Offline capability isn’t optional—it’s life-critical
  • GPUs aren’t just for models; they power real-time civic systems

* Smart cities don’t need more dashboards—they need coordination layers

What’s next for SCRAMBLE

  • Pilot deployment with a municipal corporation
  • Expand beyond floods to wildfire and earthquake intelligence
  • Scale community-led mapping for informal settlements
  • Add family reunification and accessibility-aware routing
  • Deploy city-wide systems across India—and globally

Built With

Share this project:

Updates