Unscene Gems helps users discover high-quality but overlooked places by aggregating multiple signals into a single interactive exploration experience. Instead of ranking by popularity, it surfaces locations that are underrated and contextually interesting.

Unscene Gems provides a map-based interface for exploring the world, drilling down into regions and local areas, and uncovering places that typically don't appear on mainstream recommendation platforms.

Key Innovation: Parallel AI agent orchestration with consensus-building to analyze locations from 4 different perspectives simultaneously.

• The backend (Node.js + Express) orchestrates 4 parallel AI agents via OpenRouter API
• Each agent simulates a different data source (local expertise, Reddit, TripAdvisor, news)
• Agents execute concurrently with 10-second timeouts and graceful degradation
• Results are aggregated, consensus is built, and cached for instant re-access
• The frontend (React) renders an interactive 3D globe, state panels, and detailed maps
• Real-time POI data is fetched from OpenStreetMap's Overpass API
• Frontend communicates with backend through REST APIs

Four specialized agents run in parallel using Promise.all():

Execution Flow:

User Request
    ↓
Promise.all([agent1, agent2, agent3, agent4])
    ↓
2-4 seconds → All agents complete
    ↓
Consensus Builder (sentiment analysis)
    ↓
Cached Response → Instant subsequent access

AI Model: Google Gemini Flash 2.5 (via OpenRouter)

• Temperature: 0.7
• Max tokens: 1000 per agent
• Timeout: 10s with fallback data on failure

• Express 4.18.2 - REST API server
• OpenRouter API - AI model orchestration for parallel agent execution
• axios 1.6.5 - HTTP client for external API calls
• dotenv 16.3.1 - Environment variable management

• React 18.2.0 - UI framework
• react-globe.gl 2.27.2 - 3D globe visualization (Three.js wrapper)
• Leaflet 1.9.4 - Interactive 2D maps for detailed exploration
• Three.js 0.160.0 - 3D rendering engine

• OpenRouter (https://openrouter.ai/api/v1/chat/completions) - Routes requests to Google Gemini Flash 2.5
• Overpass API (https://overpass-api.de/api/interpreter) - Real-time OpenStreetMap POI queries
• Nominatim (https://nominatim.openstreetmap.org/search) - Geocoding and global city search
• Natural Earth - Country boundary GeoJSON data
• Countries States Cities DB - Global administrative divisions CSV

• Parallel Processing: All agents execute simultaneously, not sequentially
• Graceful Degradation: Failed agents return fallback data; system continues with successful responses
• Consensus Building: Analyzes sentiment across all agent responses to determine overall confidence
• Smart Caching: In-memory Map structure caches results by {state},{country} key
• Real-time POI Search: Overpass API queries with dynamic bounding box calculation
• Global Search: Nominatim geocoding with 600ms debounce and autocomplete

• Node.js v14+
• npm v6+
• OpenRouter API key (get one here)

cd backend
npm install

Create .env file:

OPENROUTER_API_KEY=sk-or-v1-your_key_here

Run server:

node index.js

Expected output:

✅ API Key loaded: sk-or-v1-...
🕵️ WorldView Parallel Agent System running on port 3001
🚀 4 agents running in parallel: Local, Reddit, TripAdvisor, News

cd frontend/npx
npm install
npm start

Application opens at http://localhost:3000

Execute all 4 agents in parallel and return aggregated intelligence.

Response includes:

• Execution time and timestamp
• Consensus sentiment (positive/negative/neutral) with confidence level
• All 4 agent responses with source attribution
• Local spots (hidden gems, what to avoid)

Execute single agent independently.

Valid agents: local_insights, reddit_sentiment, tripadvisor_sentiment, news_sentiment

Check server status, cache size, and memory usage.

Agent Timeout Protection:

Promise.race([
  executeAgent(prompt, name),
  new Promise(resolve => setTimeout(() => 
    resolve({ success: false, error: 'Timeout' }), 10000))
])

Consensus Algorithm:

• Extracts sentiment keywords from each agent summary
• Counts positive/negative/neutral responses
• Calculates confidence based on successful agent count (high: 3+, medium: 2, low: 1)

POI Search:

• Converts radius (km) to lat/lng degrees
• Adjusts longitude for latitude-based compression: lon ± radius / cos(lat)
• Queries Overpass with dynamic bounding box
• Limits results to top 100 by synthetic rating algorithm

MIT License

Built for travelers seeking authentic, underrated experiences