Inspiration
Millions of Pakistanis receive fake NADRA, FBR, and bank messages daily. People lose money and personal data because they cannot tell what's real. We built CivicGuard to give every Pakistani citizen an AI-powered tool to instantly verify suspicious messages — in their own language.
What it does
CivicGuard lets you paste any suspicious SMS, email, or URL and get an instant AI verdict. It tells you if it's Safe, Suspicious, or Phishing, explains WHY in plain Urdu or English, shows a threat breakdown, and auto-generates a PTA complaint draft you can submit with one click.
How we built it
- Frontend: React + Tailwind CSS deployed on Vercel
- Backend: Python + FastAPI deployed on Railway
- AI Detection: HuggingFace BERT phishing classifier + URL detection model
- AI Explanation: Groq API with Llama 3 for plain language explanations
- Bilingual: Full Urdu and English support with Noto Nastaliq font
Challenges we ran into
- Combining two AI models (text + URL) into one unified confidence score
- Generating accurate Urdu explanations via Groq prompt engineering
- Cold start delays on HuggingFace Inference API free tier
- Deploying a monorepo cleanly across Vercel and Railway
Accomplishments we're proud of
- Built a fully working dual-AI pipeline in under 24 hours
- Urdu language support makes it accessible to low digital literacy users
- Auto-generated PTA complaint draft converts detection into direct action
- 98% confidence on phishing detection in live tests
What we learned
- How to combine ML models with LLM explanations in a single pipeline
- Prompt engineering for bilingual structured JSON outputs
- Full-stack deployment across Vercel and Railway with environment variables
What's next for CivicGuard
- Screenshot upload analysis
- WhatsApp message detection
- Browser extension
- Full PTA API integration
- Mobile app for iOS and Android
Log in or sign up for Devpost to join the conversation.