Lighthouse

Offline Emergency Assistant for iOS

Help when you need it. Works offline.

---

Overview

Lighthouse is an offline-first emergency assistant for anyone in a crisis — not just trained responders. Open the app, see where you are on a map, describe what's happening by voice or text, and get clear guidance on what to do next.

Under the hood it runs a local agent (on-device rules engine): it remembers incidents, geotags reports, routes SOS to the right teams, detects contradictions, and speaks responses aloud — without cloud AI.

Built for the Google DeepMind Bangalore Hackathon – Best Use of Gemma 4: Local-First Agents, and designed so anyone can use it in an emergency.

---

The Problem

During disasters such as:

• Earthquakes
• Floods
• Landslides
• Cyclones
• Building collapses
• Forest fires

internet connectivity often disappears.

Cloud AI assistants stop working. Offline chatbots only answer questions — they do not:

• remember rescue operations
• coordinate multiple incidents
• recover from missing information
• keep mission state
• continuously plan

This is where Lighthouse comes in.

---

Solution

Lighthouse is a map-first emergency app with an autonomous on-device agent.

1. Home (map) — your GPS, MapKit map, What to do guidance, quick reports, voice mic
2. Agents — SOS to ambulance, police, fire, disaster, or child-protection teams by location + country
3. Guide — full field console for power users and demos
4. Activity — message history
5. Settings — agent brain status, peer sync, reset

It runs the full agent loop locally:

Sense → Decide → Act → Verify → Recover

Offline: agent inference (rules), voice, SwiftData, SOS routing, incident memory.
Online (optional): MapKit map tiles / reverse geocoding when the network is available.

---

Core Features

Map-First Home Screen

The default tab is Home — built around where you are.

• MapKit with your location and priority-colored incident pins
• What to do card — latest agent guidance and next plan steps
• Quick report chips — "There's a fire near me", "People trapped", etc.
• Voice dock — mic, hands-free, photo + Vision OCR at the bottom of the home screen

Emergency Agents (SOS)

The Agents tab routes SOS to specialized teams based on your input, GPS, and country:

Agent	Examples
Ambulance	injured, chest pain, unconscious
Police	crime, attack, robbery
Fire & Rescue	fire, smoke, trapped
Disaster Response	flood, earthquake, storm
Child Protection	missing child, children in danger

• Auto SOS — classifies your description and dispatches
• Per-agent SOS buttons with available / assigned unit counts
• Country-aware numbers — India (112/100/101/102/1098), US/CA (911), UK (999), AU (000), else 112
• Dispatches local units (simulated on-device) and logs geotagged incidents
• Tap the number link to dial — the app does not place calls automatically

Voice-First Reporting

• Tap the mic on Home or Agents and describe what's happening
• Lighthouse responds by voice with what to do next
• Uses on-device SFSpeechRecognizer + AVSpeechSynthesizer; the rules engine parses the report

Consumer Onboarding

A 5-step flow ending with Get started — uses your GPS, no mission setup required.

1. Welcome
2. How it helps (agent loop)
3. Permissions (mic required, location recommended) + voice test
4. Optional AI / brain status
5. Ready — quick start, or advanced custom mission + presets

Presets: Earthquake Chennai, Flood Kerala, Building Collapse Mumbai.

Field Console (Guide Tab)

Power-user / demo view with:

• Agent loop bar (Sense → Decide → Act → Verify → Recover)
• Mission stats grid
• Agent briefing and plan steps
• Incident cards and recent timeline
• Play demo (scripted Chennai earthquake scenario)

Offline Mission Management

Missions, incidents, victims, resources, timeline, chat, and memory stored in SwiftData — survives app restarts and airplane mode.

Intelligent Priority Engine

Incidents scored by severity, demographics, accessibility, and age. Priority recalculates when reports change (critical / high / medium / low).

Local Error Recovery

Asks for clarification instead of guessing — e.g. "There may be another child" triggers confirm questions (age, conscious?, trapped?, injured?).

Contradiction Detection

Detects conflicting victim counts and updates the mission after confirmation.

Mission Timeline & Offline Search

Query by voice or text:

• "How many critical incidents?"
• "Where am I?"
• "Show timeline."
• "Nearest incident"

Peer Sync

Settings → Export / Import mission JSON — offline mission snapshot schema v1 for sharing between devices.

---

Design

Built for Apple’s latest system look:

• iOS 18+ Tab API, large titles, and NavigationStack
• Liquid Glass–inspired materials (.ultraThinMaterial, specular strokes, capsule voice dock)
• System blue and priority colors for incident severity
• When built with Xcode 26 / iOS 26, standard controls pick up Liquid Glass automatically

---

Tech Stack

Layer	Technology
Platform	Swift 5, iOS 18+
UI	SwiftUI (Liquid Glass–ready)
Maps	MapKit
Architecture	@Observable ViewModel → Repository → SwiftData
Database	SwiftData
Location	Core Location + CLGeocoder (country-aware SOS)
Voice	SFSpeechRecognizer + AVSpeechSynthesizer
OCR	Vision (VNRecognizeTextRequest)
Agent	On-device rules engine (parser, planner, SOS, recovery)
Project	XcodeGen

---

Project Structure

lighthouse-ios/
├── Lighthouse/
│   ├── Agent/                 # Orchestrator, parser, SOS, planner, demo
│   ├── Data/                  # MissionRepository, JSON peer sync
│   ├── Models/                # SwiftData models
│   ├── Services/              # Location, voice, field capture / OCR
│   ├── Theme/                 # Design system (glass cards, colors)
│   ├── ViewModels/            # MissionViewModel
│   ├── Views/
│   │   ├── Onboarding/
│   │   ├── Home/              # Map-first home + voice dock
│   │   ├── Agents/            # SOS screen
│   │   ├── Guide/             # Field console
│   │   ├── Activity/          # Chat log
│   │   ├── Settings/
│   │   ├── Components/
│   │   ├── MainTabView.swift
│   │   └── RootView.swift
│   ├── Assets.xcassets/
│   └── LighthouseApp.swift
├── project.yml                # XcodeGen definition
├── Lighthouse.xcodeproj/
└── README.md

---

Getting Started

Prerequisites

• Xcode 16+ (Xcode 26 recommended for Liquid Glass)
• iOS 18.0+ Simulator or device
• XcodeGen (brew install xcodegen)
• Physical device recommended for microphone / speech recognition

Build & Run

# Generate the Xcode project (after editing project.yml)
xcodegen generate

# Open in Xcode
open Lighthouse.xcodeproj

Select an iPhone simulator or device, then press Run (⌘R).

Permissions

On first launch the app requests:

Permission	Why
Microphone	Voice reports
Speech Recognition	On-device STT
Location When In Use	Geotag incidents + country SOS numbers
Camera / Photos	Field photo capture + OCR

---

Onboarding Guide

1. Welcome — tagline and feature overview
2. How it helps — Sense → Decide → Act → Verify → Recover
3. Permissions — enable mic (required), location (recommended); optional voice test
4. Optional AI — brain / variant status
5. Ready — Get started creates an emergency mission at your GPS

Advanced: custom mission name / disaster type / location, or pick a regional preset.

After Reset in Settings (clears SwiftData + captures/), onboarding appears again until a new active mission exists.

---

Agent Brain

• Lighthouse runs an offline rules engine — parse, plan, SOS, contradictions, and queries all work without the network.
• The bottom status bar shows the active brain mode.

---

Demo (Guide → Play)

Scripted walkthrough:

1. Critical collapse report (children trapped)
2. Unconscious victim update
3. Contradiction (victim count)
4. Blocked road
5. Rescue logged
6. Critical-incident query