AgentOSINT

AgentOSINT is an AI-powered OSINT investigation platform that orchestrates agents to gather, analyze, and verify intelligence in under 60 seconds, with human-in-the-loop feedback for accuracy

Comment

🎯 Inspiration

Traditional OSINT (Open Source Intelligence) investigations are time-consuming and manual. Investigators spend hours searching across multiple platforms—Reddit, Twitter, news sites, public records—with no way to verify accuracy or track sources. Information overload makes it nearly impossible to find relevant connections, and existing tools lack AI-powered analysis capabilities.

We wanted to revolutionize OSINT by combining the power of multi-agent orchestration with human intelligence, creating a system that completes comprehensive investigations in under 60 seconds while maintaining transparency and accuracy.

💡 What It Does

AgentOSINT is an AI-powered OSINT investigation platform that orchestrates five specialized agents to automate the entire intelligence gathering workflow:

  1. Collection Agent - Gathers data from Reddit, Twitter, news sources, and public records in parallel
  2. Correlation Agent - Finds connections and relationships between data points using AI
  3. Analysis Agent - Generates insights and patterns using Claude AI
  4. Validation Agent - Calculates confidence scores based on source reliability and data freshness
  5. Reporting Agent - Generates comprehensive investigation reports

Key Features:

  • ⚡ Fast: Complete investigations in 30-60 seconds
  • 🔍 Transparent: Every finding shows source, timestamp, and confidence level
  • 👤 Human-in-the-Loop: Users can verify findings, label confidence, and trigger deeper research
  • 🤖 AI-Powered: Claude AI extracts entities, generates insights, and creates reports
  • 📊 Real-Time: Live progress tracking and evidence display

🛠️ How We Built It

Architecture

  • Backend: FastAPI with async/await for parallel agent execution
  • Frontend: Vanilla HTML/JavaScript with real-time polling
  • Database: SQLite for local development (easily scalable to PostgreSQL)
  • AI: Anthropic Claude API for entity extraction, analysis, and report generation
  • Data Sources: Reddit API, Twitter API, NewsAPI, WHOIS lookups

Multi-Agent Orchestration

We built a custom orchestrator that manages agent lifecycle:

  • Agents run in parallel where possible (Collection → Correlation+Analysis → Validation)
  • Each agent logs its activities for full observability
  • Error handling ensures one agent failure doesn't break the entire investigation
  • Background tasks allow non-blocking investigation starts

Human Feedback Loop

  • Users can label evidence confidence (High/Medium/Low)
  • Users can verify evidence as accurate
  • Users can add notes and trigger additional research
  • Feedback influences future agent behavior and confidence calculations

Observability

  • Comprehensive logging to files and console
  • Health and status endpoints for monitoring
  • Request logging with timing information
  • Full error tracking with stack traces

🚧 Challenges We Ran Into

  1. Rate Limiting: Twitter API rate limits were causing 858-second delays. We fixed this by disabling automatic rate limit waiting and falling back to mock data gracefully.

  2. Form State Management: Frontend polling was resetting user input every 2 seconds. We solved this by saving and restoring form values before each update.

  3. ChromaDB Compatibility: Vector database had numpy 2.0 compatibility issues. We made it optional and added graceful fallbacks.

  4. CORS Configuration: Frontend on port 3001 couldn't connect to backend on 8000. We updated CORS to allow multiple localhost ports.

  5. Environment Variables: .env file loading wasn't working from backend directory. We fixed path resolution to load from project root.

🏆 Accomplishments We're Proud Of

✅ Complete Multi-Agent System: Successfully orchestrated 5 specialized agents working in parallel

✅ Fast Performance: Achieved 30-60 second investigation completion time

✅ User Feedback Integration: Built a unique human-in-the-loop system for accuracy verification

✅ Full Observability: Comprehensive logging and monitoring endpoints

✅ Production-Ready Code: Error handling, logging, testing, and documentation

✅ Local-First Architecture: Works completely offline with mock data fallbacks

📚 What We Learned

  • Multi-Agent Orchestration: How to coordinate multiple AI agents efficiently
  • Async/Await Patterns: Leveraging Python's async capabilities for parallel execution
  • Human-AI Collaboration: Designing effective feedback loops between users and AI systems
  • Error Resilience: Building systems that gracefully handle API failures and rate limits
  • Observability: The importance of logging and monitoring in distributed systems

🔮 What's Next

  • More Data Sources: LinkedIn, GitHub, domain registrars, public records databases
  • Advanced Correlation: Graph database for relationship mapping
  • Team Collaboration: Share investigations and findings with team members
  • Export Formats: PDF reports, CSV exports, API integrations
  • Real-Time Updates: WebSocket support for live investigation updates
  • Advanced AI: Custom fine-tuned models for specific investigation types
  • Mobile App: iOS/Android apps for on-the-go investigations

🏗️ Built With

Core Technologies

  • Python 3.12 - Backend language
  • FastAPI - Web framework
  • SQLAlchemy - Database ORM
  • Anthropic Claude API - AI reasoning and analysis
  • JavaScript/HTML/CSS - Frontend

APIs & Services

  • Anthropic Claude - AI agent reasoning
  • NewsAPI - News article collection
  • Reddit API - Social media intelligence
  • Twitter/X API - Social media intelligence
  • WHOIS - Domain registration data

Tools & Libraries

  • Uvicorn - ASGI server
  • Pydantic - Data validation
  • python-dotenv - Environment configuration
  • praw - Reddit API client
  • tweepy - Twitter API client
  • requests - HTTP client

📸 Demo

Try It Live

  1. Start backend: cd backend && python -m uvicorn main:app --reload --port 8000
  2. Open frontend: frontend/index.html in browser
  3. Enter target (person, email, username, or domain)
  4. Watch agents work in real-time!

Features to Demo

  • ⚡ Fast investigation completion (30-60 seconds)
  • 🔍 Real-time progress tracking
  • 📊 Evidence with confidence scores
  • 👤 User feedback and verification
  • 📝 Comprehensive reports

🎥 Presentation Highlights

  1. Problem: Show manual OSINT investigation taking hours
  2. Solution: Demo AgentOSINT completing investigation in 60 seconds
  3. Multi-Agent: Show agent status dashboard with parallel execution
  4. User Feedback: Demonstrate confidence labeling and research triggering
  5. Transparency: Show source tracking and confidence scores
  6. Observability: Display logs and system status

🔗 Links

  • GitHub Repository: [Your repo URL]
  • Live Demo: [Your demo URL]
  • Video Demo: [Your video URL]
  • Documentation: See README.md and OBSERVABILITY.md

👥 Team

[Your Name/Team Name]

🏅 Hackathon Track

  • **Best Use of AI/ML (Anthropic Claude)
  • **Best Multi-Agent System
  • **Most Innovative Solution
  • **Best Developer Experience

AgentOSINT - Intelligence at the Speed of Thought

Built With

Share this project:

Updates