• 

  Main search page

• 

  History page

🕵️ Deepsint

One username → one trusted profile card.
Turning scattered OSINT signals into explainable insights, fast.

🌟 Inspiration

We all hear the warnings of data breaches and the dangers of the internet.
We've also heard that someone can reconstruct everything about you using OSINT tools—without ever meeting you.

But in practice, existing OSINT tools are powerful yet slow, complex, and fragmented across 30 different tabs.
We wanted something simpler:
➡️ A single-click experience.
➡️ One username → one trusted profile card.

⚙️ What It Does

1. Input → a single username.
   
2. Crawl → Blackbird scans the open web for likely matches.
   
3. Collect → a web scraper gathers relevant, publicly available data from each hit.
   
4. Reason → Cohere embeddings correlate personas, detect behavioral patterns, and connect fuzzy signals.
   
5. Synthesize → outputs an explainable profile card:
   • Clean facts
     
   • Source links
     
   • Timestamps
     
   • Per-claim confidence scores
     
6. Guardrails → public data only, PII redaction, audit trail, and sensitive inferences disabled by default.
   

🛠️ How We Built It

Pipeline architecture:

• Discovery → Blackbird for username enumeration and candidate gathering.
  
• Extraction → site-aware scraping of bios, handles, links, timestamps.
  
• Normalization → unify fields, dedupe items, standardize time/text.
  
• Correlation → reasoning models score cross-platform matches using:
  • Handle similarity
    
  • Cross-linked bios
    
  • Writing-style cues
    
  • Semantic similarity
    
• Evidence Grading → assign confidence based on independent signals + recency.
  
• Profile Card → concise summary with sources, timestamps, and caveats.
  

🚧 Challenges We Ran Into

• Entity resolution is hard → avoiding false positives requires careful scoring & explicit caveats.
  
• Noisy & incomplete data → profiles change, vanish, or contradict each other.
  
• Anti-automation & rate limits → building a polite, robust collector without brittle hacks.
  
• UX for trust → making confidence, evidence, and caveats visible without overwhelming users.
  

🏆 Accomplishments

• Built a usable “username → trusted profile card” in minutes, not hours.
  
• Evidence-first design → every claim is traceable, timestamped, and scored.
  
• Cross-platform correlation beyond exact string matches:
  • Semantic similarity
    
  • Image reuse detection
    
• A clean, analyst-friendly UI → facts first, exploration second.
  

📚 What We Learned

• In OSINT, speed is nothing without explainability.
  
• Confidence scores + links build trust—and catch mistakes early.
  
• Most value comes from normalization & correlation, not just bigger models.
  
• Ethical defaults are essential for adoption and long-term viability.
  

🚀 What’s Next for Deepsint

• Name-based discovery → privacy-respecting search by name to widen correlation.
  
• Image-based discovery → profile picture correlation layered with AI reasoning & embeddings.
  

Built With

• cohere
• python
• sqlite
• streamlit