ScamShield AI - Agentic Honey-Pot

Overview

ScamShield AI is an agentic honey-pot platform for scam detection and intelligence extraction. Online scams — bank fraud, UPI fraud, phishing, and fake investment offers — adapt based on user responses, making static detection ineffective.

This system detects scam intent, autonomously engages scammers with a believable human persona, handles multi-turn conversations, extracts actionable intelligence, and reports structured results to evaluation endpoints.

Live API Docs: honeypot-production-b319.up.railway.app/docs · GitHub: Tusharkanta407/HoneyPot

Problem Statement

Scammers use urgency tactics, payment redirection, and evolving scripts. Traditional rule-based filters miss adaptive fraud. ScamShield runs an AI agent that wastes scammer time, gathers evidence, and surfaces risk signals — UPI IDs, phone numbers, phishing links, and suspicious keywords — without revealing it is a honeypot.

High-Level Architecture

Incoming Message
      ↓
Session Manager (sessionId)
      ↓
Scam Detection Agent
      ↓
Persona Strategy Agent
      ↓
Multi-Turn Engagement Loop
      ↓
Intelligence Extraction
      ↓
Termination Check
      ↓
Final Result Callback

Key Features

Scam Detection & Autonomous Engagement

• Detects scam or fraudulent messages on first contact
• Activates an autonomous AI agent with a believable human persona
• Never reveals detection to the scammer
• Adapts responses dynamically across conversation turns

Multi-Turn API

• POST /honeypot — one message per request with full session history
• API-key secured access (x-api-key header)
• Structured JSON request/response format

Intelligence Extraction

Extracted silently from conversation text:

• Bank account numbers
• UPI IDs
• Phone numbers
• Phishing links
• Suspicious keywords

Dashboard Experience

• Live conversation view with scammer vs agent transcript
• Extracted intelligence panel (payment methods, amounts, keywords)
• Real-time scam analysis with risk score and scam type classification
• Session metadata — timing, locale, channel

API Example

Request:

{
  "sessionId": "wertyu-dfghj-ertyui",
  "message": {
    "sender": "scammer",
    "text": "Your bank account will be blocked today. Verify immediately.",
    "timestamp": 1770005528731
  },
  "conversationHistory": [],
  "metadata": {
    "channel": "SMS",
    "language": "English",
    "locale": "IN"
  }
}

Response:

{
  "status": "success",
  "reply": "Why is my account being suspended?"
}

Tech Stack

| Layer | Technology | |-------|------------| | Backend | FastAPI, Python 3.11+ | | AI Agents | Scam detector, persona, extraction agents | | Auth | API key (x-api-key) | | Deploy | Docker, Railway | | Docs | OpenAPI / Swagger at /docs |

Project Structure

app/
├── main.py                 # FastAPI entry point
├── api/honeypot_handler.py # Main controller
├── agents/
│   ├── scam_detector_agent.py
│   ├── persona_agent.py
│   └── extraction_agent.py
├── tools/                  # Detection & extraction tools
├── memory/session_store.py
├── callbacks/guvi_callback.py
└── schemas.py

What I Learned

• Designing multi-agent pipelines where detection, engagement, and extraction run as separate coordinated steps
• Session-based conversation memory for stateful multi-turn REST APIs
• Building production FastAPI services with Docker deployment to Railway
• Balancing believable persona behavior with reliable structured intelligence output

Impact

• Autonomous honeypot that turns scam messages into actionable fraud intelligence
• Multi-turn engagement increases data captured per session
• Production-ready API with documented endpoints for integration and demos
• Built for real-world scam patterns common in India (UPI, bank blocks, investment fraud)

Links

• API Documentation: https://honeypot-production-b319.up.railway.app/docs
• Repository: https://github.com/Tusharkanta407/HoneyPot