| Slide 1: Title |
| INDIA AI IMPACT BUILDATHON |
| Challenge 2: Agentic Honeypot for Scam Detection |
| Team: ScamShield AI |
| Slide 2: Introduction |
| Project Name: ScamShield AI |
| Tagline: An autonomous agentic honeypot system that detects scam messages, engages scammers with believable AI personas, and extracts actionable intelligence (UPI, Bank Accounts, Links). |
| Slide 3: The Problem |
| 1) THE PROBLEM |
| 🚨 What is happening? |
| 500,000+ scam calls/msgs daily |
| ₹60+ crore daily losses |
| 47% Indians affected |
| Target: UPI, Loans, KYC |
| ⚠️ Why is it a problem? |
| Passive detection fails |
| Scammers evolve fast |
| No intelligence gathered |
| Citizens feel helpless |
| 👥 Who is affected? |
| Elderly & Non-tech savvy |
| Middle-class families |
| Banks (Reputation loss) |
| Law Enforcement (Overload) |
| Slide 4: Our Solution |
| 2) OUR SOLUTION |
| 🎯 Detection |
| IndicBERT + Rules |
| Hindi, English, Hinglish |
| 90%+ Accuracy Target |
| 🤖 Engagement |
| LangGraph Agent |
| Dynamic Personas (Elderly, Eager) |
| Up to 20 turns conversation |
| 🔍 Extraction |
| Extracts UPI, Bank, IFSC |
| Captures Phishing Links |
| Structured JSON Output |
| 🚀 Integration |
| REST API Endpoint |
| Mock Scammer API Ready |
| Dockerized & Scalable |
| Slide 5: How It Works |
| 3) HOW IT WORKS |
| STEP 1: INPUT |
| Message arrives: "You won 10 lakh! Send OTP." |
| STEP 2: DECISION (AI) |
| Confidence > 70%? Yes -> Trigger Honeypot |
| STEP 3: ENGAGE |
| Agent (Persona): "Oh wow! How do I get it?" |
| STEP 4: EXTRACT & OUTPUT |
| Scammer shares UPI -> We capture it -> JSON Report |
| Simple Flow: Message In ➡️ Scam Check ➡️ Fake Victim Talks ➡️ Proof Extracted |
| Slide 6: Proof It Works |
| 4) PROOF IT WORKS |
| ✅ Live API Demo: POST /honeypot/engage works in <2s |
| 📊 Test Metrics: Validated on 100+ scam samples |
| 🇮🇳 Real Context: Handles Hindi/Hinglish perfectly |
| 🛡️ Extraction: Success: Captured UPIs & Links |
| Slide 7: A Nuance We Handled |
| 5) A NUANCE WE HANDLED |
| 🌐 Mixed Language |
| Handles Code-Mixing (Hindi + English) |
| IndicBERT trained on Indian datasets |
| 😇 Over-Polite Scams |
| Personas don't get suspicious |
| Elderly/Confused tones matches scammer's pace |
| 🔄 Persistence |
| Redis Session State |
| Remembers Context |
| Engages for 10+ turns |
| Slide 8: Trade-Offs & Failure Cases |
| 6) TRADE-OFF & 7) FAILURE CASE |
| ⚖️ Trade-Off: Depth vs Speed |
| Choice: Maximize engagement depth |
| Why: More turns = More Intelligence |
| Cost: Higher latency (1-2s) |
| Note: Not optimized for instant blocking |
| ❌ Failure Case: Short Convos |
| Struggle: Scammer stops after 1 msg |
| Result: No intelligence extracted |
| Mitigation: 'Eager' persona to bait |
| Struggle: Novel templates (Low Conf) |
| Slide 9: Submission Details |
| SUBMISSION DETAILS |
| Team: ScamShield AI |
| Challenge 2: Agentic Honeypot |
| Tech Stack: IndicBERT, LangGraph, Groq, FastAPI |
| Contact: missionupskillindia@hclguvi.com |
| Subject: ScamShield AI PPT || India AI Impact Buildathon |
| Deadline: 13th Feb 2026 |
