AI Guardian: Local AI Security

Protect your data with local AI – no cloud, no leaks, just privacy.

Comment 6

Inspiration

Every day, millions of people accidentally share sensitive information online – bank details, OTPs, passwords, and personal data. Grandparents fall for phishing emails, small businesses leak customer data, and teenagers share too much on social media. I wanted to build something that protects everyone, especially those who aren't tech-savvy.

What it does

I built AI Guardian as a privacy-first solution using 100% local AI – no cloud, no data leaks, no API costs.

How we built it

  • Backend: Python + FastAPI for lightning-fast API endpoints
  • Frontend: Streamlit for beautiful, interactive dashboard
  • AI: Ollama + Llama 3.2 (runs completely locally!)
  • Visualization: Plotly for stunning risk gauges and analytics
  • Deployment: GitHub Codespaces (fully reproducible)

The architecture is simple but powerful:

  1. User pastes text in the dashboard
  2. FastAPI sends it to local Llama 3.2 model
  3. AI analyzes for risks (phishing, leaks, PII)
  4. Returns risk score + safe rewrite + security tips
  5. Dashboard visualizes everything beautifully

Challenges we ran into

The biggest challenge was making the local LLM fast enough for real-time use. I solved this by:

  • Implementing model caching (loads once, stays in memory)
  • Using smaller, optimized models (Llama 3.2 3B)
  • Adding smart timeout handling
  • Creating fallback keyword detection when AI is warming up

Another challenge was making the dashboard beautiful AND functional. I spent hours on CSS gradients, animations, and the perfect color scheme (green = safe, orange = suspicious, red = dangerous).

Accomplishments that we're proud of

Every day, millions of people accidentally share sensitive information online – bank details, OTPs, passwords, and personal data. Grandparents fall for phishing emails, small businesses leak customer data, and teenagers share too much on social media. I wanted to build something that protects everyone, especially those who aren't tech-savvy.

I was inspired by stories of people losing their life savings to phishing scams and friends accidentally sharing passwords in Discord chats. The problem isn't that people are careless – it's that we don't have simple tools to check messages BEFORE sending them. AI Guardian solves this with a beautiful, easy-to-use dashboard that runs completely offline.

What we learned

This project was an incredible learning journey across multiple domains:

          Technical Skills
  • FastAPI : Building production-ready REST APIs with automatic documentation
  • Streamlit : Creating beautiful, interactive dashboards with custom CSS
  • Ollama Integration : Running local LLMs without cloud dependencies
  • Plotly: Designing custom gauges, charts, and data visualizations
  • Error Handling : Managing timeouts, connection issues, and fallback systems

  • Session State : Tracking user history and analytics in Streamlit

      AI/ML Skills

  • Prompt Engineering : Crafting prompts that get structured JSON responses

  • Model Caching : Loading models once to avoid repeated startup delays

  • Local LLMs : Understanding the trade-offs between cloud and local AI

  • JSON Parsing : Extracting structured data from free-text AI responses

  • Fallback Systems : Using regex when AI is warming up or slow

    Cybersecurity Knowledge

  • Phishing Patterns : Recognizing common scam templates and urgent language

  • PII Detection : Identifying emails, phones, bank details, and SSNs

  • OTP Risks : Understanding why one-time passwords are dangerous to share

  • Social Engineering : How attackers manipulate people into sharing data

  • Privacy by Design : Building systems that never send data to the cloud

    Design & UX

  • Color Psychology : Green for safe, orange for suspicious, red for dangerous

  • Glass-morphism : Creating modern, translucent UI elements

  • Responsive Design : Making dashboards work on different screen sizes

  • User Flow: Guiding users from pasting text to understanding results

Project Management

  • Hackathon Planning: Delivering a complete project under deadline
  • Solo Development: Managing all aspects from backend to frontend
  • Documentation: Writing clear READMEs and submission materials
  • Time Management: Prioritizing features that matter most to judges

Key Takeaways

  1. Local AI is the future of privacy – No data leaves your machine
  2. Beautiful UI matters – Judges and users judge by what they see first
  3. Fallbacks are essential – Always have a plan B when AI is slow
  4. Small models can do big things – Llama 3.2 3B is powerful enough
  5. Hackathons teach more than tutorials – Building real projects is the best learning

What's next for AI Guardian: Local AI Security

Phase 1: Browser Extension (Immediate - 2 Weeks)

  • Chrome/Firefox Extension that scans emails and messages in real-time
  • Right-click Context Menu – "Check with AI Guardian" for any selected text
  • Gmail/Outlook Integration – Scan emails before sending
  • Social Media Protection – Warn before posting sensitive info on Twitter, Facebook

Built With

Share this project:

Updates