π Project Story β FlipRedact π
Inspiration β¨
Have you ever wanted to use powerful Large Language Models (LLMs) but couldn't because your prompts and outputs were full of sensitive data? During our internships, we constantly ran into this problem. Manually handling PII (Personally Identifiable Information) was not only time-consuming but also incredibly risky. That's why we were inspired to create FlipRedact, an automatic privacy layer for generative AI.
What It Does π‘οΈ
FlipRedact acts as a secure middleman, automatically detecting and redacting sensitive data like names, IDs, and emails before your prompts ever touch an LLM. When you get the response back, you can safely unredact the data with a single click. We've made privacy and convenience a priority!
Key Features π‘
- You're in Control: Choose exactly what you want to be redacted.
- Smart Redaction: Redact multiple instances of the same PII at once.
- Filter with Ease: Filter redactions by the type of PII.
- Transparency: See a confidence score for each detected PII.
- Personalized View: Toggle between dark mode and light mode.
How We Built It π οΈ
We built this solution from the ground up to be both fast and accurate. Hereβs how we did it:
- We started with a lightweight NLP model and fine-tuned it using custom regex rules for high-speed PII detection.
- For even greater accuracy, we integrated powerful Hugging Face NER models (Named Entity Recognition).
- We created a robust redaction/unredaction layer to act as a secure middleware between you and LLM APIs.
- Finally, we developed a sleek web interface to showcase the seamless user experience of FlipRedact.
Challenges We Conquered π§
Building this project wasn't easy! As full-time interns, we navigated some tough challenges:
- Night Owls: We could only work on this late at night after our workdays.
- Remote-First: We coordinated our entire project online across different cities, from Singapore to Chongqing.
- Juggling Act: Balancing a demanding hackathon schedule with our full-time internship commitments was a real test of our time management skills!
Accomplishments We're Proud Of π
Despite the challenges, we accomplished some amazing things:
- We built a fully functional end-to-end system in less than 72 hours!
- We successfully combined cutting-edge machine learning with rule-based methods to achieve the perfect balance of accuracy and efficiency.
- We designed a privacy solution that doesn't compromise on usability.
What's Next for FlipRedact π
Our journey is just beginning! We have some big plans to make FlipRedact even better:
- Smarter Detection: Further fine-tune our model for even more precise PII detection.
- Expand Horizons: Go beyond PII and add redaction for new categories like medical or financial data.
- More Access: Package FlipRedact as a browser extension or a mobile SDK so everyone can use it.
- Deeper Protection: Explore advanced Privacy-Enhancing Technologies (PETs), like homomorphic encryption.
- Community Driven: We plan to open-source FlipRedact to empower other developers to build privacy-aware applications.
Built With
- css
- fastapi
- hugging-face
- javascript
- python
- pytorch
- react


Log in or sign up for Devpost to join the conversation.