Flood warnings don't matter if you can't access them. Emergency services are useless if the network is down. When a 300-year flood hit Thailand, a developer named Flamehaven01 watched infrastructure collapse and realised something: disaster response tools are built assuming connectivity. That assumption fails exactly when you need them most.
So they built Flood Ready - a progressive web app that runs a 1.5-billion-parameter AI model entirely offline, with three layers of fallback resilience and a peer-to-peer QR code relay system for when absolutely everything goes dark.
What Makes This Different
Most disaster apps are just mobile-friendly websites. They look good, load fast on WiFi, and become completely useless the moment connectivity drops. Flood Ready is offline-first - it assumes the network is gone and builds from there.
The AI model runs locally on your device using WebGPU and WebAssembly. No server calls. No API keys. No dependency on external infrastructure. You can ask it questions, get safety guidance, assess flood risk, and plan evacuation routes without a single byte leaving your phone.
The system has three fallback tiers. First, it tries the full AI model if your device can handle it. If that's too heavy, it drops to a smaller quantized version. If that still doesn't work, it falls back to a rule-based expert system - simple decision trees, no AI, but still functional.
And if your device is completely offline but someone nearby has connectivity? The app generates QR codes containing compressed requests. Another user can scan it, relay the query through their connection, and pass back a QR code response. It's absurdly low-tech and brilliantly practical.
The Technical Reality
Running a 1.5B parameter model on a phone is not trivial. It requires aggressive model quantization - reducing precision to shrink file size without destroying accuracy. It means careful memory management so the app doesn't crash mid-inference. And it means accepting that performance will vary wildly depending on device capability.
But here's what matters: it works. On a mid-range Android phone, you get usable responses in a few seconds. On older hardware, it's slower but functional. And if your hardware truly can't handle it, the fallback tiers ensure you still get help.
The offline-first architecture means all critical data is cached locally. Maps, evacuation routes, emergency contact lists, medical guidance - it's all stored in the service worker. Updates sync when connectivity returns, but the core functionality never depends on it.
Why This Approach Matters
Disaster scenarios are edge cases by definition, and edge cases are where most software breaks. We build for the common case - fast networks, reliable power, predictable infrastructure. When those assumptions collapse, so do our tools.
Flood Ready inverts that thinking. It's built for the worst case. Connectivity is a bonus, not a requirement. The AI is helpful but not essential. The interface works on battered old phones because that's what people have in emergencies.
There's a broader lesson here about resilience in software design. Every system has failure modes. The question is whether you've planned for them or just hoped they won't happen. Progressive enhancement isn't just about browser compatibility - it's about building systems that degrade gracefully when conditions deteriorate.
What Could This Unlock
Flood Ready is one person's response to one disaster, but the pattern is replicable. Offline-first PWAs with on-device AI could work for earthquake response, wildfire evacuation, power grid failures, medical emergencies in remote areas - any situation where infrastructure is unreliable or absent.
The QR relay system is particularly clever. It treats humans as network nodes. If one person has connectivity, everyone nearby can benefit through physical proximity and QR codes. It's peer-to-peer networking without requiring any special protocols or permissions.
This won't replace traditional emergency services. But it fills a gap - the hours or days when official systems are overwhelmed and people need immediate, local information to make decisions. That window is where this kind of tool could genuinely save lives.
Building for disasters means building for reality. Connectivity fails. Power fails. Infrastructure fails. The tools that matter most are the ones that keep working anyway.