Back to Projects
AI Systems
FraudShield
Most scam detectors fail because they only see one signal.
FraudShield analyzes text, links, images, and voice together to detect scams in real time with clear, explainable outputs.
2025
2 Months
Active
Problem
Most scam detection systems operate on a single input type, making them ineffective against multi-format attacks.
Solution
Built a real-time AI pipeline for scam detection across text, URL, image, and voice inputs with explainable outputs.
Impact
- 93.7% detection accuracy
- <2s latency
- Supports 10+ Indian languages
- Multi-modal input pipeline (text, URL, image, voice)
How it was built
- Built multi-modal fraud detection pipelines
- Designed real-time FastAPI inference system
- Integrated external threat intelligence APIs
- Implemented multilingual NLP using MuRIL
- Developed explainable verdict engine
- Deployed production-ready system
Tech Stack
Frontend
Next.js, TypeScript, Tailwind CSS
Backend
FastAPI, Node.js, Express
AI / ML
MuRIL (HuggingFace), scikit-learn, Tesseract OCR, Librosa
Database
Supabase (PostgreSQL)
Security APIs
Google Safe Browsing, VirusTotal, PhishTank, WHOIS
Deployment
Vercel, Railway
Challenges
Handling noisy real-world inputs (SMS, OCR, voice)
Built normalization and preprocessing pipelines
Balancing accuracy with real-time latency
Optimized async processing and caching