Open-source AI platform for disaster information integration
English | 日本語
When disasters strike, fragmented information systems create a "dark 72 hours" where no authority has a unified picture of evacuee numbers, shelter capacities, or relief supply distribution. DDCR solves this by normalizing heterogeneous disaster data into a unified Common Operating Picture (COP) — without replacing any existing system.
- "AI surfaces uncertainty, humans decide." — AI does not generate "truth." It structures what is known, what is unknown, and where data conflicts exist.
- Vendor ecosystem, not vendor replacement. — DDCR works with existing disaster system vendors through co-developed adapters.
- Hybrid architecture. — Sensitive personal data stays on-premise; anonymized analytics leverage cloud AI (Google Gemini, Vertex AI).
- Provenance first. — Every data point carries lineage metadata: who reported it, when, how, and with what confidence.
- Open standards only. — JSON Schema, GeoJSON, W3C PROV-O. No proprietary middleware.
Shows real-time data flowing from diverse sources with different confidence levels:
- Government Systems (high confidence) — automated CSV/DB ingestion
- Staff Reports (medium-high) — field observations, Excel uploads
- NPO/Partner Orgs (medium) — structured report forms
- SNS/Public Info (low) — AI-extracted, requires verification
Existing Systems → Adapter Layer → AI Normalization → Decision Support → COP Dashboard
(untouched) (vendor co-built) (schema mapping, (conflict flagging, (what we know,
provenance tracking) natural language what we don't,
queries) who said what)
| Component | Environment | Rationale |
|---|---|---|
| Personal data processing | On-premise | Legal requirement; internet-independent |
| Real-time COP generation | On-premise | Must operate during connectivity loss |
| Schema mapping rule generation | Google Cloud (Gemini) | Anonymized data only |
| Model evaluation | Google Cloud (Vertex AI) | Anonymized test datasets |
| Satellite damage assessment | Google Cloud (Geospatial AI) | Public imagery |
DDCR defines open schemas for disaster data entities. All schemas are JSON Schema compliant and designed to be used with any database or API — no proprietary middleware required.
| Schema | Description |
|---|---|
shelter.schema.json |
Emergency shelter with capacity, evacuees, supply status, and vulnerable population tracking |
supply.schema.json |
Relief supply tracking with allocation and provenance |
decision.schema.json |
Immutable decision audit trail — who decided what, based on which evidence |
Provenance tracking — Every observed value includes:
- Who reported it (source entity)
- How it was collected (field observation, system input, AI normalized, etc.)
- Confidence level (high / medium / low)
- When it was observed
Conflict preservation — When multiple sources report different values for the same field, DDCR preserves all values with their provenance rather than silently overwriting:
{
"evacuees": {
"value": 147,
"observed_at": "2026-01-01T06:30:00Z",
"provenance": {
"reported_by": "JGSDF_3rd_Regiment",
"method": "field_observation",
"confidence": "high"
},
"conflicts": [
{
"source": "Yokohama_City_System",
"value": 203,
"observed_at": "2026-01-01T05:00:00Z",
"confidence": "medium"
}
]
}
}Decision-makers see: "JGSDF reports 147 (field observation, 06:30, high confidence) — conflicts with city system showing 203 (05:00, medium confidence)."
| Layer | Technology | License |
|---|---|---|
| Database | PostgreSQL + PostGIS | PostgreSQL License |
| On-premise LLM | Ollama (open models) | MIT |
| API Server | FastAPI (Python) | MIT |
| Authentication | Keycloak | Apache 2.0 |
| Geospatial | Geolonia Maps SDK | MIT |
| Cloud AI | Google Gemini API, Vertex AI | — |
DDCR was born from the lessons of the 2024 Noto Peninsula earthquake, where 34,000+ displaced people across 600+ shelters could not be tracked by any single system. Building a victim database required integrating 15 incompatible data sources — a challenge that recurs in every major disaster in Japan.
Two multi-stakeholder workshops have validated the problem framing and technical approach:
- Ishikawa Prefecture (March 2025) — 50 participants including prefectural government, 2 disaster-affected municipalities, and 12+ companies
- Tokushima Prefecture (November 2025) — 50 participants including prefectural/municipal government and 10+ disaster-related organizations
Both workshops confirmed: the diversity of disaster data, the critical need for cross-organizational databases, and that monolithic systems alone are insufficient.
- Code for Japan — Applicant. Civic tech nonprofit (est. 2013). Operates data interoperability platforms for municipalities (Saga City, Hamamatsu City). Built Tokyo's COVID-19 dashboard (replicated to 65+ government sites, Good Design Gold Award 2020).
- DIT/CC — Strategic partner. Permanent secretariat of D-CERT (Digital Agency's official disaster digital support team). Founded by LINE Yahoo, NTT, SoftBank, PwC, Tokio Marine Holdings, Fujifilm, and Mitsui Sumitomo Insurance.
- Kanagawa Prefecture — Government partner (9.4M residents, 33 municipalities). DIT/CC's director serves as the prefecture's Chief Digital Officer.
DDCR is currently in the schema definition and prototyping phase. This repository will grow as development progresses. Contributions and feedback are welcome.
MIT License. See LICENSE.
Building open infrastructure for disaster resilience.