ScamShield — AI-Powered Digital Scam & Phishing Guard

Inspiration

Every day, millions of digital users—especially senior citizens and first-time smartphone adopters—receive fraudulent WhatsApp messages, fake UPI refund requests, SMS phishing alerts, and phishing emails. Most victims fall prey not because they lack intelligence, but because fraudsters exploit urgency, fear, and technical jargon.

We built ScamShield to act as an instant, empathetic AI digital safety companion that doesn't just say "SCAM" or "SAFE", but explains why a message is dangerous in simple, relatable terms and provides clear step-by-step guidance on what to do next.

How We Built It

ScamShield is engineered as a modern full-stack web application powered by Google Gemini 3.6 Flash and TypeScript:

• Multimodal AI Analysis Engine: Users can paste raw text messages, suspicious URLs, or directly upload screenshots of WhatsApp chats and SMS alerts.
• Vision OCR & Signal Extraction: Using Gemini's vision capabilities, ScamShield extracts text snippets from images and flags specific high-risk phrases.
• Mathematical Risk Scoring Model: Computes a normalized scam probability score $S \in [0, 100]$ based on weighted risk vectors.
• Multilingual Vernacular Explanations: Translates complex security breakdowns into regional languages (Malayalam, Hindi, Tamil, and English) so non-English speakers can easily protect themselves.
• Official Emergency Response: Integrates direct reporting workflows and instant links to the official Indian Cyber Crime Helpline (1930) and cybercrime.gov.in.

Scam Risk Calculation Formula

The AI risk score $S$ is computed using a weighted Bayesian signal model evaluating $N$ distinct scam indicators:

$$S = \min\left(100, \sum_{i=1}^{N} w_i \cdot \sigma(f_i)\right)$$

Where:

• $w_i$ represents the severity weight for signal $i$ (e.g., OTP request $w_{OTP} = 35$, fake UPI link $w_{UPI} = 30$, urgency pressure $w_{urgency} = 20$).
• $\sigma(f_i)$ is the activation state of red flag $f_i \in {0, 1}$.
• High scores ($S > 60$) automatically trigger the High-Risk Official Reporting Hotline (1930) modal.

Challenges We Faced

1. Handling Image OCR & Formatting Noise: Real-world WhatsApp screenshots contain dark mode backgrounds, emojis, and low-contrast text. We tuned system prompts with Gemini 3.6 Flash to extract exact text snippets accurately.
2. Preventing User Panic: Balancing urgent security warnings with reassuring, step-by-step action plans without creating unnecessary anxiety.
3. Iframe & Clipboard Security Constraints: Supporting smooth clipboard pasting and pre-formatted report sharing across restricted browser preview sandboxes by building graceful fallback mechanisms.

What We Learned

• Empathy in Security Design: Clear visual gauges, progress meters, and regional language translations dramatically improve user trust compared to technical diagnostic codes.
• Interactive Reinforcement: Adding interactive mini-quizzes and bite-sized safety tips helps users retain cybersecurity awareness long after analyzing a message.

Built With

• ai
• cloudrun
• css
• express.js
• freebuff
• icons
• lucide
• node.js
• react
• tailwind
• typescript
• vite