SUBMISSION_MANIFEST.md

🛡️ Sentinel Scam Honeypot - GUVI Submission

Quick Start

# Install dependencies
pip install -r requirements.txt

# Set API key
export GROQ_API_KEY="your-key-here"

# Run server
uvicorn app.main:app --host 0.0.0.0 --port 7860

API Endpoint

POST /api/guvi/analyze
Header: x-api-key: GUVI_HACKATHON_V2

Features

Feature	Status
Scam Detection	✅ Regex + LLM
Multi-turn Conversations	✅ Session-based
Human-like Personas	✅ 5 Indian personas
Intelligence Extraction	✅ Regex-primary (UPI, Phone, Bank, etc.)
GUVI Callback	✅ Automatic on completion
Key Rotation	✅ Multi-key support
Static Fallbacks	✅ Budget exhaustion safety
Rate Limit Handling	✅ Proactive cooldowns

Architecture

app/
├── main.py              # FastAPI entry point
├── agents/
│   ├── orchestrator.py  # Main pipeline
│   ├── persona_engine.py # Human-like responses
│   └── scam_detector.py  # Detection logic
├── core/
│   ├── llm_client.py    # Groq with rotation
│   └── personas.py      # Persona definitions
└── utils/
    ├── guvi_handler.py  # GUVI API adapter
    └── extractors.py    # Regex intelligence

Personas

• Sharma Uncle (65, Delhi) - Elderly, trusting
• Rahul Kumar (24, Bihar) - Desperate jobseeker
• Meena Patel (45, Gujarat) - Worried housewife
• Shyam Sharma (52) - Scared citizen
• Curious Investor (35) - Analytical, cautious

Response Format

{
  "status": "success",
  "scamDetected": true,
  "scamConfidence": 0.95,
  "reply": "Arre yaar, UPI se kaise bhejun?",
  "extractedIntelligence": {
    "upiIds": ["scammer@upi"],
    "phoneNumbers": ["+919876543210"]
  },
  "agentNotes": "[HIGH RISK] Banking fraud detected"
}

GUVI Callback

Automatically sent when:

• High-value intel captured (UPI/Bank) + 2 turns
• Medium intel + 3 turns
• 6 turn hard cap
• Budget exhausted

---

Team: AvinashAnalytics
Project: Sentinel Scam Honeypot