Not a replacement for paper — another option alongside it.
ShelterAI is a fully offline iPhone app that turns an everyday device into a disaster shelter registration tool. Point the camera at any Japanese ID card; Gemma 4 E2B, running entirely on-device, extracts the evacuee's information in seconds. No internet, no subscriptions, no proprietary hardware.
Tracks: Main | Global Resilience | Cactus (Local-First Mobile)
- Auto-shutter camera — Real-time card detection via VNDetectRectanglesRequest; auto-captures after 1.4s of stable detection
- Gemma 4 multimodal analysis — Image + prompt → structured JSON (name, address, DOB, card type)
- Vision OCR fallback — Apple Vision as cross-check; confidence-gated hybrid merge
- Perspective correction — CIPerspectiveCorrection + contrast boost for dim gymnasium lighting
- 5 card types — My Number Card, driver's license, passport (MRZ), residence card, health insurance card
- Manual input — 15-region injury body-part picker, special-needs toggles, gender/age pickers
- Family group inference — Same surname + address block → automatic grouping suggestion
- Consent gate — Explicit privacy consent before every save
- Rate limiting — 5 registrations/minute cap to prevent volunteer mistakes
- Voice announcements — Japanese TTS confirmations for hands-free use
- GPS stamping — Location tagged on each record (emergency mode)
- 5 deterministic presets — Injury triage, supply estimates, special-needs roster, summary, foreign resident support. Generated by Swift code from the database — 100% accurate, instant response
- Free-form Gemma 4 Q&A — General shelter operations knowledge (hygiene, supplies, procedures). Streaming generation via CoreMLLLM
- Conversation management — Reset button, session-scoped memory
Why intelligent routing? A 2B model excels at flexible tasks (multimodal extraction, open-ended Q&A) but struggles with structured data retrieval. Presets go to code; free questions go to AI. Each query reaches the system best suited to answer it.
- Real-time statistics — Age, gender, injury severity, special needs breakdown
- Supply estimation — 1–30 day slider based on Cabinet Office guidelines (water 3L/person/day, food, blankets, diapers, formula, sanitary products, medical kits)
- Foreign resident detection — Name-based heuristics (ASCII/katakana-only names)
- SQLite (GRDB) — 5 schema migrations, Codable records
- Search + filter —
.searchablebar + attribute chip filters - Individual deletion — Per-record delete with confirmation
- CSV export — UTF-8 BOM for Excel. Offline collection → handoff to municipal officials via AirDrop/USB/email once communications restore
- Distribution tracking — Per-evacuee supply checklists with progress bars
- Sample data — 18 realistic evacuees for testing and demos
- Zero network calls in the entire codebase
- No analytics, no telemetry, no Firebase
- Two dependencies total (GRDB.swift, CoreML-LLM)
- Test-mode data auto-deleted on app termination
Camera → CardPreprocessor → ┬─ Gemma 4 E2B (CoreML) ─┐
└─ Vision OCR (fallback) ──┤
├→ Confidence merge → Register
│
AI Assistant ─── Preset? ─── YES → ShelterAnalytics (Swift code, 100%)
└── NO → Gemma 4 E2B (streaming, best-effort)
| Layer | Technology |
|---|---|
| UI | SwiftUI (iOS 18), TabView, NavigationStack, .searchable |
| AI | CoreMLLLM (Gemma 4 E2B), chunked CoreML model |
| Vision | VNRecognizeTextRequest, VNDetectRectanglesRequest, CIPerspectiveCorrection |
| Camera | AVCaptureSession + AVCaptureVideoDataOutput (real-time detection) |
| Database | GRDB.swift (SQLite), 5 migrations |
| Audio | AVSpeechSynthesizer (ja-JP) |
| Location | CoreLocation (emergency mode) |
- Xcode 16.0+ / iOS 18.0+ device
- ~3 GB free disk space for model files
- Neural Engine required for Gemma 4 inference (simulator uses Vision OCR fallback only)
git clone https://github.com/Resolver-TNG/ShelterAI.git
cd ShelterAI
open AnpiApp.xcodeprojModel files (~2.7 GB) are not included in the repository:
pip install huggingface_hub
python -c "
from huggingface_hub import snapshot_download
snapshot_download(
repo_id='mlboydaisuke/gemma-4-E2B-coreml',
local_dir='Models/gemma-4-E2B-coreml'
)
"Accept the Gemma 4 license on Hugging Face before downloading.
Select your device → Set Development Team → ▶ Run. SPM resolves GRDB and CoreML-LLM automatically. First model load takes 10–30 seconds.
Competition: Gemma 4 Good Hackathon (deadline May 18, 2026)
Japan's Digital Agency proved that digital ID processing reduces shelter intake time by 90% — but existing solutions require internet, paid subscriptions, or proprietary hardware. ShelterAI removes every barrier by bringing Gemma 4 directly onto the device.
At a 500-person shelter: paper takes ~12.5 hours. ShelterAI takes ~0.8 hours. 11+ hours freed for care, distribution, and coordination.
ShelterAI is released under Apache 2.0 because disaster preparedness belongs in the commons. Fork it, adapt it, deploy it — no permission needed, no purchase order required.
Built for the shelters that have no internet, no budget, and no time — where a phone with Gemma 4 is the only extra tool that fits in a volunteer's pocket.
Apache License 2.0. Gemma 4 model weights are subject to Gemma Terms of Use.