FraudGuard

Inspiration

India lost over ₹11,000 crore to cyber fraud in 2023. The victims aren't careless — they're shopkeepers handed a fake PhonePe screenshot, students receiving fake KYC calls, families targeted by fake RBI officers. The problem isn't awareness. It's that in the moment of decision, there's no tool to verify. We built FraudGuard to close that gap.

What it does

FraudGuard analyses suspicious UPI payment screenshots, scam messages, phishing links, and recorded phone calls using Gemini 2.0 Flash and returns an instant fraud verdict — with exact red flags, a confidence score, and clear action steps in Hindi or English. No app install. No technical knowledge required.

How we built it

Single-file web app — no frameworks, no backend. The Gemini API is called directly from the browser using three modalities: multimodal vision for screenshot analysis, text reasoning with Google Search grounding for messages and links, and native audio input for call recordings. We engineered a domain-specific system prompt with deep knowledge of NPCI transaction ID formats, legitimate UPI app UI signatures, and India-specific scam patterns so Gemini reasons like a fraud investigator, not a generic AI assistant.

Challenges we ran into

Getting Gemini to return consistent structured JSON across all three input modalities required significant prompt engineering — especially for blurry or partially visible screenshots. Tuning the confidence thresholds so the system doesn't over-flag legitimate transactions was equally critical; a tool that cries wolf is useless.

What we learned

Gemini's multimodal vision goes far beyond OCR — it performs genuine visual reasoning, detecting pixel-level artifacts from image editing that no text-based model could catch. Combining that with real-time Google Search grounding creates a fraud detection layer that's both technically deep and practically accessible.