ScamBait AI

ScamBait AI is an agentic honeypot system that engages scammers, extracts intelligence, and disrupts their operations by wasting their most valuable resource - time.

We deployed ScamBait AI across two real-world channels to demonstrate its capabilities.

First: Telegram. Telegram is widely used by scammers. We deployed a bot that behaves like a real human. Every message from the scammer is forwarded to our backend, where an LLM generates realistic replies. This architecture allows us to reuse the same detection and extraction logic across hundreds of simultaneous scam conversations - without rebuilding the system's core intelligence.

Second: Phone-based scams. Using Twilio, we host real phone numbers that connect directly to our system. When a scammer calls, our model generates responses in real time, and ElevenLabs converts them into natural, confused human-like voices. To the scammer, it feels completely real.

---

Quick Links

Demo video, live dashboard, Telegram bot, and source code.

Resource	Link
Telegram Bot	https://web.telegram.org/k/#@honeytrap_ai_bot
Public GitHub	github.com/diyaavirmani/SCAM-BAIT
API	honey-api-wr74.onrender.com
Bot (Webhook)	honey-bot-9kce.onrender.com
** Dashboard **	https://scam-bait.onrender.com/

Architecture Diagram

High-level system architecture (ScamBait AI).

---

How It Works

End-to-end flow: from scammer contact to intelligence extraction.

---

Working of the Telegram Bot

Screenshot showing the Telegram bot in action: scammer message and AI persona reply.

---

What This Project Does

Overview of features: persona, detection, extraction, dashboard, voice.

1. Multi-channel victim persona The same "Persona" (confused elderly) persona responds via Telegram bot and (optionally) Twilio voice. Judges or testers send messages as a "scammer"; the system replies in character.

2. Scam detection Rule-based + ML (TF-IDF + SVM) pipeline. Classifies into types: DIGITAL_ARREST, UPI_SCAM, JOB_SCAM, SEXTORTION, etc.

3. Intelligence extraction Regex + normalization to pull out: phone numbers, UPI IDs, emails, APK links, phishing links, crypto wallets, social handles, IFSC codes, and suspicious keywords.

4. Concurrency and reliability Session locking and semaphore (e.g. 30 concurrent) so many simultaneous sessions don't mix or crash. Graceful fallbacks on timeout/errors.

5. **Voice ** Twilio to Deepgram (STT) to LLM to ElevenLabs (TTS) to back to call. Barge-in and silence handling.

---

Tech Stack

Backend, database, voice, bot, frontend, and deployment.

Layer	Technologies
Backend	Python 3.11+, FastAPI, LangGraph, LangChain (Groq / Cerebras), scikit-learn
Database	SQLite,SQLAlchemy
Voice	Twilio (Programmable Voice), Deepgram (STT), ElevenLabs (TTS)
Bot	python-telegram-bot, aiohttp (webhook server)
Frontend	React 18, TypeScript, Vite, TailwindCSS, Recharts, Framer Motion, Lucide React
Deploy	Render (API, Bot, Dashboard)

---

Repo Structure

Main folders and files: app, bot, dashboard, evaluation.

scam_bait_bot_tts_web/
├── app/                    # FastAPI backend
│   ├── agents/             # detection, extraction, persona, timeline, hallucination_filter
│   ├── services/           # audio_orchestrator (voice)
│   ├── workflow/           # LangGraph graph
│   ├── main.py             # API, /api/v1/honeypot, /ws/dashboard, /api/v1/stats
│   ├── voice_router.py     # Twilio /voice/incoming, /voice/stream
│   ├── websockets.py       # Dashboard WebSocket manager
│   ├── database.py         # SessionManager (SQLAlchemy)
│   └── models.py           # Pydantic request/response models
├── bot/                    # Telegram bot
│   ├── bot_service.py      # Handlers, session management, API calls to honeypot
│   └── bot_config.py       # Token, API URL, API key
├── dashboard/              # React frontend (Vite + TS)
├── evaluation/             # Reddit scam dataset, expand_prompts, run_evaluation
├── run.py                  # Start API (uvicorn)
├── run_bot.py              # Start bot (webhook or polling)
├── render.yaml             # Render blueprint (honey-api, honey-bot, honey-dashboard)
├── requirements.txt        # Backend + voice deps
└── bot/requirements.txt    # Telegram bot deps

---

Run Locally

Clone, set environment variables, then start API, bot, and dashboard.

1. Clone and install

   git clone https://github.com/diyaavirmani/SCAM-BAIT
   cd scam_bait_bot_tts_web
   pip install -r requirements.txt
   pip install -r bot/requirements.txt

2. Environment

   • Copy .env.example to .env.
   • Set: API_KEY, GROQ_API_KEY (or CEREBRAS_API_KEY), TELEGRAM_BOT_TOKEN, DATABASE_URL (or leave unset for SQLite).
   • For voice: TWILIO_*, DEEPGRAM_API_KEY, ELEVENLABS_API_KEY, ELEVENLABS_VOICE_ID.

3. Start API

   python run.py

   API: http://localhost:8000, docs: http://localhost:8000/docs.

4. Start Telegram bot

   python run_bot.py

   Without RENDER_EXTERNAL_URL / PRODUCTION_BOT_URL it runs in polling mode.

Pre-existing Tools and Assets Used

Disclosure of third-party APIs, libraries, and assets.

We disclose the following third-party services, libraries, and assets:

• APIs and services

  • Groq and/or Cerebras - LLM inference for persona and detection fallback.

  • OpenAI-compatible (e.g. LangChain) - used only if configured.

  • Deepgram - speech-to-text for voice calls.

  • ElevenLabs - text-to-speech for voice responses.

  • Twilio - telephony and media streams for voice.

  • Render - hosting for API, bot, and frontend.

  • Telegram - Bot API and (optionally) BotFather-created bot.

• Libraries

  • FastAPI, Uvicorn - backend.
  • LangChain, LangGraph - orchestration and LLM calls.
  • scikit-learn - TF-IDF + LinearSVC for scam detection.
  • python-telegram-bot, aiohttp - Telegram bot and webhook server.
  • React, Vite, TypeScript, TailwindCSS, Recharts, Framer Motion, Lucide React - frontend.
  • Pydantic, SQLAlchemy, psycopg2-binary - data and DB.

• Assets

  • No proprietary artwork or media was used; UI is custom CSS/Tailwind and Lucide icons.
  • Evaluation prompts are derived from a Reddit scams-style dataset (see evaluation/dataset/); no external licensed datasets.

• Code / design

  • Project and PRDs were developed internally; no pre-existing "scam honeypot" codebase was copied.
  • Architecture and prompts were designed for this hackathon/project.

Vibe Log - Prompts and Workflow

How we designed prompts and the main workflow (persona, detection, extraction, graph, voice).

Persona

• Goal: Sound like a confused, elderly person who might fall for scams (so scammers keep talking) but never give real money or data.
• Prompt style: System prompt defines "Meena" - forgetful, hard of hearing, needs things repeated, asks "who is this?" and "is this the bank?". Replies are short, slightly tangential, and in character. We avoid any tone that sounds like a bot or security researcher.
• Iteration: We tuned length and confusion level so the model doesn't refuse to engage or break character. Hallucination filter (in app/agents/hallucination_filter.py) blocks responses that leak "I'm an AI" or safety refusals.

Detection

• Goal: High recall on scammy messages without blocking normal SMS (OTP, delivery, etc.).
• Approach:
  • Rules first: Keyword sets and legit senders (e.g. "amazon", "flipkart", "order #"). If rule score >= threshold then SCAM.
  • ML second: TF-IDF + SVM on a small labeled set; used when rules are unsure.
  • Scam type: Same keyword sets mapped to DIGITAL_ARREST, UPI_SCAM, JOB_SCAM, SEXTORTION for dashboard and analytics.
• Prompt (when using LLM fallback): "Is this message a scam? Reply SCAM or NOT_SCAM and optionally type." We kept it minimal to avoid over-refusal.

Extraction

• Goal: Pull phones, UPI, emails, APK links, crypto, etc., even when obfuscated ("at" to "@", "dot" to ".", spaced digits).
• Workflow: Normalize text (in app/agents/extraction.py) then run regex for each type. Merge results from raw + normalized. No LLM for extraction in the main path.

Graph (LangGraph)

• Flow: load_session to detection (every turn) to extraction (on full history) to persona (LLM in character) to validate_persona_output to timeline (confidence/summary) to persist and optional callback.
• Decisions: "Completed" when confidence is high and we decide to end the conversation (e.g. "Have a great day" or equivalent). Otherwise we stay in "engaging" and keep responding.

Voice

• Goal: Same persona over the phone. Twilio streams audio to Deepgram to text to same LLM persona to ElevenLabs TTS to stream back.
• Prompts: Reuse the same Meena system prompt; no separate voice-specific prompt. We added silence handling ("Hello? Are you there?") and barge-in (stop TTS when user speaks).

What we'd do next

• More persona variants (e.g. "greedy lottery victim") and A/B test which gets more intel.
• Richer extraction (e.g. LLM for messy text) and export to CSV/JSON for analysts.
• Dashboard: filters by scam type, export, and optional auth.

---

License

MIT License. See LICENSE. Copyright (c) 2026 team ARCANE.

---

Thanks