DeadZone

An offline-first AI-powered disaster response network.

When networks go down, DeadZone keeps rescue operations moving.

Built for Smart India Hackathon 2026.

---

The problem

When a flood, landslide or cyclone takes down cellular and internet in a region, victims cannot call for help and responders cannot coordinate — precisely when speed saves lives. Existing emergency apps assume connectivity. DeadZone assumes its absence.

What it does

COMMUNICATE → UNDERSTAND → PRIORITISE → COORDINATE → DELIVER

Stage	Capability
Communicate	SOS reaches a responder with no cellular and no internet, over a field command node's own Wi-Fi hotspot
Understand	Local LLM explains weather and disaster risk in plain language
Prioritise	Deterministic START triage scores every incoming SOS
Coordinate	Responders see a live offline Bengaluru map and assign nearest teams
Deliver	Everything syncs to a central node the moment a link returns

Plus a coverage meter that reports how long since the device last heard from the command node — a measured liveness signal, never an estimated range.

Transport, precisely. The app is a web page running in the phone's browser. There is no Bluetooth, no Nearby Connections, no RSSI and no device-to-device relay — a browser exposes none of those. Each phone holds one WebSocket to the responder's laptop over its hotspot. See "Transport reality" in CLAUDE.md.

---

Prerequisites

• Node 20+
• pnpm 9+ (corepack enable)
• CouchDB 3.3+ — runs natively as a Windows service on the dev laptop; infra/docker-compose.yml is kept for machines with a working Docker daemon
• Ollama on the machine acting as edge node
• A phone with a modern browser. No app install, no Android Studio, no JDK — it is a web app.

---

Setup

pnpm install
pnpm typecheck

# CouchDB: native service, or compose where Docker works
#   docker compose -f infra/docker-compose.yml up -d
curl http://localhost:5984/            # must answer, not hang

# Seed the local database
pnpm seed

# Edge node: pull the model and serve on the LAN
ollama pull phi4-mini
OLLAMA_HOST=0.0.0.0 OLLAMA_ORIGINS=* OLLAMA_KEEP_ALIVE=-1 ollama serve

Copy .env.example to .env and fill in local values. Never commit .env.

---

Workspace

apps/
  mobile/       Vite + React web app — civilian + responder
  api/          Express coordination, sync, LAN mesh
  ai/           WeatherGPT service (edge node)
  dashboard/    offline tile server — basemap, glyphs, sprites
packages/
  schema/       Zod contracts — source of truth
  comms/        envelope, peer table, coverage state
  triage/       START rules + greedy allocation
  ui/           placeholder — design tokens
data/
  seed/         fictional demo records
  tiles/        offline Bengaluru vector tiles + glyphs + sprites
infra/          docker-compose (CouchDB)
docs/           architecture, decisions, progress, runbook

---

Commands

Command	Does
pnpm typecheck	Recursive typecheck across all packages
pnpm test	Unit tests
pnpm seed	Load seed data into local PouchDB
pnpm sync-test	Interrupt/resume replication test (cleans up after itself)
pnpm dev:api	Express + WebSocket on 0.0.0.0
pnpm dev:ai	WeatherGPT service

---

Running the offline demo

1. Start CouchDB, the tile server and the AI service on the edge-node laptop
2. Enable the laptop hotspot; connect the phone to it
3. Turn the phone's mobile data off — visibly, so a judge can see it
4. Open http://<laptop-IP>:8443 on the phone and send an SOS
5. It appears on the responder's screen with no cellular and no internet
6. Re-enable connectivity and watch replication drain to the central node

Full choreography in docs/demo-runbook.md.

---

Architecture in one line

Everything writes locally first. The network is treated as an optional enhancement, never a dependency. See docs/ARCHITECTURE.md.

---

Documentation

File	Contents
CLAUDE.md	Standing context: stack, settled decisions, cut list, guardrails. Read first.
docs/ARCHITECTURE.md	System design and rationale
docs/PROGRESS.md	Current status, per track
docs/demo-runbook.md	The ninety-second demo, with a proven/unproven status table
docs/pitch-transport-coverage.md	Approved wording for transport and coverage, and the claims that must not be made
docs/decisions.md	Cut list and settled decisions
docs/GOALS.md	What success looks like
docs/build_logs.md	Build failures and their fixes

---

apps/mobile needs an explicit task before you touch it

It originated as a Figma Make export and may still be edited there, so unannounced changes can be overwritten or conflict. Do not modify anything under apps/mobile without an explicit task to do so.

---

Status

Pre-alpha. Built to a six-day sprint deadline for SIH internal evaluation. Not production software — see the cut list in docs/decisions.md for what is deliberately unimplemented.

If you are modifying this for production use in a disaster-response context, read the critical notes in CONTRIBUTING.md first.

---

Licensing

This project is licensed under the MIT License. See LICENSE for the full text.

By contributing, you agree your work is licensed under the same terms. See CONTRIBUTING.md for details.

Using in production

This is prototype code. Before deploying to a real disaster-response system:

• Audit the cut list in docs/decisions.md — several items (encryption, conflict resolution, transport reliability) are critical for production
• Have triage and allocation logic reviewed by someone with disaster-management or medical training
• Verify against your region's actual protocols (this uses Indian IMD thresholds)
• Test exhaustively on real hardware and networks
• Implement proper error handling and logging
• Set up monitoring and alerting