AI-powered scam-baiter — an AI persona answers scammer calls, keeps them on the line, transcribes everything, mines the funniest moments, and auto-generates short-form social clips.
Yale SOM MGMT 673 — Generative AI & Social Media | Team Dead Ringers
ak3289 · ms4852 · na628 · prl28 · vm522 · yl2854
[TODO: screenshot — call table with scam badges, "Simulate New Call" button]
[TODO: screenshot — modal with textarea pre-filled with IRS scam lines, spinning loader]
[TODO: screenshot — chat bubbles (scammer red / Betty green), highlight sidebar with virality bars, clip player]
[TODO: screenshot — Grandma Betty card with backstory, voice ID, scam-type badges]
| Component | State | What the mock does | What the real version needs |
|---|---|---|---|
| Classifier Agent | Mock | Always returns {is_scam: true, confidence: 0.87, scam_type: "irs_impersonation"} |
Anthropic Claude call with TeleAntiFraud-28k prompt |
| Persona Agent | Mock | Always upserts and returns "Grandma Betty" | Per-scam-type persona selection from DB, dynamic generation |
| Dialogue Agent | Partial | Cycles through fixtures/betty_responses.json (13 lines); Claude SDK wired, toggled by MOCK_CLAUDE=true |
Set MOCK_CLAUDE=false + ANTHROPIC_API_KEY |
| Highlight Miner Agent | Mock | Returns 3 template highlights with realistic scores | Sentiment + volume spike analysis on real transcript |
| Editor Agent | Stub | Runs ffmpeg to generate a 1-second black placeholder .mp4; falls back to empty file if ffmpeg missing |
Real ffmpeg assembly from audio segments + caption burn-in |
| Social Agent | Mock | Hardcoded 8 hashtags + caption from top highlight snippet | Claude call for caption; post-time from Kitboga virality model |
| Twilio Webhooks | Stub | Accepts POST /voice/incoming, returns TwiML <Connect><Stream>; WebSocket handler scaffolded |
Live Twilio number + ngrok tunnel |
| Deepgram STT | Mock | WebSocket receives audio frames but returns no transcript | DEEPGRAM_API_KEY + live audio stream from Twilio |
| ElevenLabs TTS | Mock | Returns silent bytes | ELEVENLABS_API_KEY + voice streaming |
/api/calls/simulate |
Fully working | Runs the entire 6-agent pipeline from scripted utterances — no Twilio required | Used for all demos |
TL;DR for the demo: set
MOCK_CLAUDE=falseand supplyANTHROPIC_API_KEYto get real Claude dialogue. Everything else runs in mock mode with no other keys required.
- Python 3.11+
- Node 20+ and npm 10+
ffmpegon PATH (optional — Editor Agent falls back to a stub if missing)
git clone <repo-url>
cd scamslayer# With uv (recommended — faster, no activation needed)
pip install uv
uv sync
# Or plain pip
python3 -m venv .venv && source .venv/bin/activate
pip install -e ".[dev]"cd frontend && npm install && cd ..cp .env.example .env
# Minimum to run the simulate demo:
# ANTHROPIC_API_KEY=sk-ant-... (only needed if MOCK_CLAUDE=false)
# MOCK_CLAUDE=true (default — skips all LLM calls)uvicorn backend.app.main:app --reload --port 8000The database (scamslayer.db) is created automatically on first boot.
cd frontend && npm run dev
# → http://localhost:5173Open http://localhost:5173, click Simulate New Call, paste scammer lines, and watch the pipeline run. No phone or API key required.
# Install and authenticate ngrok
ngrok http 8000
# Copy the https URL → NGROK_URL in .env
# Point your Twilio number's Voice webhook to: <NGROK_URL>/voice/incomingSee docs/SETUP.md for full API-key acquisition steps.
# All backend tests (92 passing as of Phase 4)
.venv/bin/pytest backend/tests/ -v
# Single file
.venv/bin/pytest backend/tests/test_simulate.py -vscamslayer/
├── backend/
│ ├── app/
│ │ ├── main.py FastAPI entrypoint + lifespan
│ │ ├── config.py pydantic-settings; all env vars
│ │ ├── routes/
│ │ │ ├── calls.py POST /simulate · GET /calls · GET /calls/{id} · …
│ │ │ ├── personas.py CRUD /personas
│ │ │ ├── clips.py /clips · /clips/{id}/generate
│ │ │ └── voice.py Twilio /voice/incoming + WebSocket /voice/stream
│ │ ├── agents/
│ │ │ ├── classifier.py ClassifierAgent → ClassifierResult
│ │ │ ├── persona.py PersonaAgent → PersonaResult
│ │ │ ├── dialogue.py DialogueAgent → DialogueResult
│ │ │ ├── highlights.py HighlightMinerAgent → HighlightsResult
│ │ │ ├── editor.py EditorAgent → EditorResult
│ │ │ ├── social.py SocialAgent → SocialResult
│ │ │ └── fixtures/
│ │ │ └── betty_responses.json 13 in-character stall lines
│ │ └── db/
│ │ ├── models.py SQLAlchemy 2.x: Call · Persona · TranscriptSegment · …
│ │ └── session.py async engine · get_db · init_db
│ └── tests/ pytest-asyncio, 92 tests
├── frontend/
│ └── src/
│ ├── pages/
│ │ ├── Dashboard.tsx call table + simulate modal
│ │ ├── CallDetail.tsx chat bubbles + highlight sidebar + clip player
│ │ ├── Personas.tsx read-only persona cards
│ │ └── Clips.tsx clip grid with inline video
│ └── lib/api.ts typed fetch helpers for all endpoints
├── docs/
│ ├── ARCHITECTURE.md mermaid agentic flow diagram
│ ├── SETUP.md API key acquisition walkthrough
│ ├── ETHICS.md legal & ethical constraints
│ └── PRESENTATION.md slide outline
├── analysis/ Jupyter notebooks (virality, classifier)
├── .env.example all env vars with placeholder values
├── docker-compose.yml Redis for arq background jobs
└── pyproject.toml Python deps + pytest config
See docs/ARCHITECTURE.md for the full agentic flow diagram.
See docs/ETHICS.md. Read before touching any call-related code.
See docs/SETUP.md for where to get every credential.