Built at the Gemini 3 Singapore Hackathon 2025 | Powered by Google Gemini 2.0 Flash
Global scam operations stole over $1 trillion in 2024 (GASA). Deepfake and video call fraud surged 1,500% across Asia-Pacific. In Singapore alone, government impersonation scams caused S$240M+ in losses from 3,400+ cases.
Today's scammers don't just send phishing texts. They run organized call centers with fake police uniforms, staged government offices, and counterfeit badges, calling victims live on video, accusing them of crimes, and demanding immediate money transfers under threat of arrest.
Every anti-scam tool on the market today works after the damage is done: blocking known numbers, flagging suspicious texts, or filing reports post-incident. Nothing protects victims during the call itself.
ScamBreak AI is the first real-time AI intervention system for live video call scams.
The moment a suspicious video call begins, a swarm of four Gemini-powered AI agents analyze it simultaneously:
| Agent | Role |
|---|---|
| 🔍 Vision Sentinel | Scans video feed for fake badges, green screens, staged backdrops, and counterfeit documents |
| 🧠 PsyOps Analyzer | Detects psychological manipulation tactics: fear, urgency, isolation, and coercion patterns |
| 📋 Protocol Validator | Checks if the caller violates official government communication protocols |
| ⚡ Strike Dispatcher | Aggregates all signals and makes the final verdict: scam or safe |
When a scam is detected, ScamBreak delivers short, direct warning cards designed for victims under stress. No paragraphs. No jargon. Just clear, actionable messages:
- "Fake badge detected"
- "They're manipulating you"
- "Hang up now. It's a scam."
With a single click, ScamBreak auto-generates a complete incident report containing:
- 📸 AI-captured evidence screenshots
- 🔎 Visual identity search (Google Cloud Vision + Gemini)
- 📞 Phone number risk assessment
- 📝 Full transcript analysis with flagged phrases
- 🤖 AI-powered risk profiling and threat assessment
The report is ready to submit to law enforcement authorities immediately.
Live Video Call
│
▼
┌─────────────────────────┐
│ Frame Capture Engine │ Captures video frames every 12s
└────────────┬────────────┘
│
▼
┌─────────────────────────┐
│ Gemini 2.0 Flash │ Multi-agent swarm analysis
│ ┌───────┬───────┐ │
│ │Vision │PsyOps │ │
│ ├───────┼───────┤ │
│ │Proto │Strike │ │
│ └───────┴───────┘ │
└────────────┬────────────┘
│
┌──────┴──────┐
▼ ▼
┌──────────┐ ┌──────────────┐
│ Alert UI │ │ Vision API │
│ (Cards) │ │ Face Search │
└──────────┘ └──────┬───────┘
▼
┌──────────────┐
│ Auto Report │
│ Generator │
└──────────────┘
- AI/ML: Google Gemini 2.0 Flash (multi-agent swarm), Google Cloud Vision API
- Frontend: Next.js 16, React, TypeScript
- Styling: Tailwind CSS with custom glassmorphism design system
- APIs: Real-time video frame analysis, reverse image search, transcript analysis
# Clone the repository
git clone https://github.com/hanselhansel/scam-protector.git
cd scam-protector
# Install dependencies
npm install
# Set up environment variables
cp .env.example .env.local
# Add your GEMINI_API_KEY to .env.local
# Run the development server
npm run devOpen http://localhost:3000 to see ScamBreak AI in action.
| Metric | Scale |
|---|---|
| Global scam losses (2024) | $1.03 trillion |
| Video call fraud growth (APAC) | +1,500% |
| Singapore impersonation scam losses | S$240M+ |
| Time to detect scam with ScamBreak | < 30 seconds |
ScamBreak doesn't detect scams after they happen. It breaks them while they're happening.
MIT