AI-powered maritime surveillance that catches illegal fishing vessels before they disappear.

IUU (Illegal, Unreported, and Unregulated) fishing is a $23 billion criminal enterprise destroying ocean ecosystems. Evidence exists across disconnected databases — satellite AIS tracking, marine protected area boundaries, RFMO registries, vessel ownership records, sanctions lists — but no single agency monitors them all.

DarkFleet consolidates 7 data sources into a unified risk-scoring dashboard, ranks vessels by suspicion level, and uses Claude AI to generate real-time intelligence briefs for investigators.

Built at the Claude Hackathon @ Imperial College London

flowchart LR
    GFW[Global Fishing Watch API] -->|vessels & events| Ingest[Ingest Pipeline]
    WDPA[WDPA MPA Boundaries] --> DB[(PostgreSQL + PostGIS)]
    RFMO[RFMO Registries] --> DB
    Sanctions[OpenSanctions] --> Scoring
    Ingest --> DB
    DB --> Scoring[Risk Scoring Engine]
    Scoring -->|7 weighted signals| DB
    DB --> API[FastAPI Backend]
    API --> Frontend[Globe Dashboard]
    API --> Claude[Claude AI Analyst]
    Claude -->|intelligence briefs & chat| Frontend

Every vessel gets a 0–100 risk score computed from 7 signals over a rolling 12-month window:

Alert tiers: 🔴 Red ≥ 80 · 🟠 Amber ≥ 60 · ⚪ Clear < 60

Hard triggers override the score upward: IUU-blacklisted vessels are forced to Red, recently detained vessels to Amber.

Claude acts as a maritime intelligence analyst, receiving full vessel context — risk signals, event history, ownership data, flag history — and producing:

• Intelligence briefs: Structured assessment with key risk factors, pattern analysis, and recommended actions
• Interactive chat: Investigators can ask follow-up questions with streaming responses via SSE
• Contextual awareness: Claude understands RFMO regulations, flag state implications, and IUU behavioral patterns

sequenceDiagram
    participant Investigator
    participant Frontend
    participant API
    participant Claude

    Investigator->>Frontend: Select suspicious vessel
    Frontend->>API: GET /vessels/{mmsi}/signals
    Frontend->>API: POST /ai/brief/{mmsi}
    API->>API: Gather vessel context (events, signals, ownership)
    API->>Claude: Send context + analyst system prompt
    Claude-->>API: Structured intelligence brief
    API-->>Frontend: Brief + risk breakdown
    Investigator->>Frontend: Ask follow-up question
    Frontend->>API: POST /ai/chat (streaming)
    API->>Claude: Conversation + vessel context
    Claude-->>Frontend: SSE streamed response

Select a high-risk vessel to get a Claude-generated intelligence assessment with key risk factors and recommended actions.

Drill into individual risk signals: loitering events, RFMO status, flag changes, and more.

Claude generates a structured risk assessment and answers follow-up questions about enforcement actions, jurisdiction, and next steps.

Generate downloadable incident reports with vessel profile, RFMO status, event timeline, and score breakdown.

• Docker & Docker Compose
• API keys: Global Fishing Watch, Mapbox, Anthropic

# Clone and configure
git clone https://github.com/your-org/darkfleet.git
cd darkfleet
cp .env.example .env
# Edit .env with your API keys: GFW_API_KEY, MAPBOX_TOKEN, ANTHROPIC_API_KEY

# Launch
docker-compose up

# Open http://localhost:8000

Click "Reload Ships" in the dashboard to ingest vessels from Global Fishing Watch. The scoring engine runs automatically after ingest.

# Marine Protected Area boundaries (WDPA GeoJSON)
python scripts/wdpa_ingest.py path/to/wdpa.geojson

# RFMO authorisation lists
python scripts/rfmo_ingest.py --wcpfc wcpfc.csv --iccat iccat.csv

# Vessel ownership enrichment (GISIS)
curl -X POST http://localhost:8000/enrich/all

darkfleet/
├── app/
│   ├── main.py              # FastAPI app, CORS, router registration
│   ├── models.py            # Vessel, Event, MPAZone, RFMOAuthorised, VesselOwnership
│   ├── scoring.py           # 7-signal risk scoring engine
│   ├── database.py          # Async SQLAlchemy + PostGIS setup
│   └── routers/
│       ├── ai.py            # Claude intelligence briefs & streaming chat
│       ├── scoring.py       # /alerts, /score endpoints
│       ├── vessels.py       # Vessel queries & trails
│       ├── reports.py       # PDF/HTML incident reports
│       ├── ingest.py        # GFW data import trigger
│       ├── mpa.py           # MPA zone GeoJSON
│       └── enrich.py        # GISIS ownership enrichment
├── frontend/
│   └── index.html           # Single-page dashboard (Mapbox + AI chat)
├── scripts/
│   ├── gfw_ingest.py        # Global Fishing Watch vessel & event ingestion
│   ├── wdpa_ingest.py       # MPA boundary loader
│   ├── rfmo_ingest.py       # RFMO registry loader
│   └── imo_gisis_enrich.py  # IMO/GISIS ownership lookup
├── docker-compose.yml
├── Dockerfile
└── requirements.txt

• Fragmented data: Most maritime databases lack public APIs. GISIS requires scraping, IUU lists come as PDFs/Excel, RFMO formats vary by organization.
• Spatial query performance: Optimizing PostGIS spatial joins (ST_DWithin, ST_Within) to avoid N+1 explosions across hundreds of vessels and thousands of events.
• Real-time AI context: Building rich vessel context documents for Claude that include scoring signals, event timelines, and ownership chains — all assembled on-the-fly from multiple database queries.

Built with Claude at Imperial College London.

MIT