AI-Powered Geolocation Intelligence from a Single Photo
Geointer is an intelligent geolocation system built with the GitHub Copilot SDK. Upload any photograph of a building or place, and Geointer will identify its exact location on a map — typically within 2 minutes.
No GPS metadata needed. Works with screenshots, social media images, and photos with stripped EXIF data.
Geointer runs a fully automated multi-phase pipeline:
- EXIF & Metadata Extraction — Parses GPS coordinates, camera info, and timestamps. If GPS is present, location is resolved instantly.
- LLM Visual Analysis — A large language model analyzes the image for visual clues (architecture, signs, vegetation, road markings) and estimates an approximate location.
- Reverse Image Search — The photo is sent to a visual search engine to find matching images across the web.
- LLM Image Comparison — Each discovered image is compared with the original using a vision model. Only confirmed matches (same building/place) proceed further.
- LLM Location Extraction — For matched pages, a language model extracts street addresses formatted for geocoding.
- Geocoding & Consensus — Extracted addresses are geocoded via Nominatim (OpenStreetMap). A quorum-based consensus algorithm selects the best location when multiple sources agree.
Results stream to the UI in real-time via Server-Sent Events (SSE).
┌──────────────┐
│ Upload Photo│
└──────┬───────┘
│
┌───────────────┼───────────────┐
│ │ │
▼ ▼ ▼
┌─────────────┐ ┌───────────┐ ┌──────────────┐
│ EXIF Parse │ │ LLM Geo │ │ Lens Upload │
│ GPS, camera,│ │ Visual │ │ Reverse image│
│ timestamps │ │ analysis │ │ search │
└──────┬──────┘ └─────┬─────┘ └──────┬───────┘
│ │ │
│ ┌────┘ ▼
│ │ ┌──────────────────┐
│ │ │ Lens Search │
│ │ │ Find matching │
│ │ │ pages online │
│ │ └────────┬─────────┘
│ │ │
│ │ ▼
│ │ ┌──────────────────┐
│ │ │ Page Analysis │
│ │ │ │
│ │ │ Visit pages │
│ │ │ ▼ │
│ │ │ LLM image compare│
│ │ │ (match score %) │
│ │ │ ▼ │
│ │ │ LLM address │
│ │ │ extraction │
│ │ │ ▼ │
│ │ │ Geocode per page │
│ │ └────────┬─────────┘
│ │ │
▼ ▼ ▼
┌─────────────────────────────────────────────┐
│ Consensus Scoring │
│ Cluster results within 50km, quorum vote │
│ Most agreeing sources → best location │
└──────────────────────┬──────────────────────┘
│
▼
┌──────────────────┐
│ Map + Results │
│ Multi-marker │
│ map with all │
│ candidates │
└──────────────────┘
External Services:
┌──────────────┐ ┌──────────────┐ ┌──────────────┐
│ GitHub │ │ Apify │ │ Nominatim │
│ Copilot SDK │ │ Visual Search│ │ OpenStreetMap│
└──────────────┘ └──────────────┘ └──────────────┘
- GitHub Copilot SDK — AI model orchestration
- Next.js 15 — Full-stack React framework
- Leaflet / react-leaflet — Interactive map with multi-marker support
- Nominatim (OpenStreetMap) — Forward and reverse geocoding
- Apify — Visual search integration
- Server-Sent Events — Real-time streaming pipeline results
- Multi-marker map — All geocoded candidates shown with confidence-based coloring
- Image comparison — Vision model confirms matches with percentage confidence scores
- Consensus scoring — Quorum-based algorithm: locations with the most agreeing sources win
- Preliminary results — LLM geolocation provides a fast estimate while deeper analysis runs
- Evidence panel — Per-page analysis cards showing match status, extracted addresses, and geocoded coordinates
- Camera direction — Estimates compass bearing from visual cues
- Node.js 18+
- npm
# Clone the repository
git clone https://github.com/kloba/geointer.git
cd geointer/web
# Install dependencies
npm install
# Configure environment
cp .env.example .env.local
# Edit .env.local with your API keys
# Start development server
npm run devOpen http://localhost:3000 and upload a photo.
| Variable | Description |
|---|---|
APIFY_API_TOKEN |
Apify API token for visual search |
GITHUB_TOKEN |
GitHub token with Copilot access |
geointer/
├── src/
│ └── agent/
│ └── tools/ # Custom tool implementations
│ ├── extract-exif.ts
│ ├── llm-compare.ts
│ ├── llm-extract-location.ts
│ ├── llm-geolocate.ts
│ ├── forward-geocode.ts
│ ├── reverse-geocode.ts
│ ├── sun-position.ts
│ └── visit-page.ts
├── web/
│ ├── app/
│ │ ├── page.tsx # Main page
│ │ ├── api/analyze/ # Pipeline API endpoint
│ │ └── globals.css
│ ├── components/
│ │ ├── MapView.tsx # Multi-marker Leaflet map
│ │ └── ResultsPanel.tsx # Analysis results UI
│ └── lib/
│ ├── pipeline.ts # Orchestration pipeline
│ └── types.ts # Type definitions
├── docs/
│ ├── README.md # Full documentation (problem, solution, setup, RAI)
│ └── spec/
│ └── geointer-agent.md
├── presentations/
│ └── Geointer.pptx # Challenge presentation
├── AGENTS.md # Agent pipeline & tool docs
├── mcp.json # MCP server config
└── README.md
Full documentation: See
docs/README.mdfor detailed setup, deployment, architecture, and responsible AI notes.
- OSINT & Threat Intelligence — Identify locations from surveillance or social media imagery
- Real Estate Verification — Confirm property locations from listing photos
- Insurance Claim Validation — Verify incident locations from submitted photos
- Journalism & Fact-Checking — Geolocate images for story verification
- Law Enforcement — Support investigations with photo-based location identification
This project demonstrates how the GitHub Copilot SDK can replace a complex backend stack with a single orchestration layer. The SDK provides:
- Model access via Copilot subscription (no separate AI service billing)
- Tool orchestration — custom tools for geocoding, web scraping, image analysis
- Streaming — real-time SSE events for progressive result delivery
Adding new capabilities is as simple as defining new tools — no infrastructure changes needed.
MIT
