Neutrafy

What Inspired Me

The inspiration for Neutrafy came from personal observations and experiences:

US Elections affected news-led polarization in every major country of the world
A shocking statistic that 80% of people form their opinions based solely on headlines
My growing concern about the increasing polarization in society
Personal experience with family members growing apart due to different news consumption habits
Read research on how subtle language changes can dramatically impact perception
The realization that while we can't control the news, we can control how it's presented to us

What I Learned

I did not expect a local model to be of this calibre. While it definitely works like a smaller model (more prompt tuning needed), it definitely beats all other smaller models on the quality of outputs. I was shocked when it just worked zero-shot for outputting headlines. Then, after extensive testing, I got it to a stable state working for most of the headlines.

This project has been fun and rewarding:

Technical Skills:

Chrome's local AI API and it's limitations
Learned how to wrangle with prompts specifically for google nano for consistent results

UX Insights:
- Sometimes the best interface is the simplest
- Visual feedback is crucial for building user trust
- Users prefer immediate results over perfect results (from my user research on this specific idea)

How I Built It

The development process followed these stages:

Research & Planning:

Studied existing news headline patterns to look for emotional trigger words
Conducted user interviews about news consumption habits with close friends and family

Design Phase:

Created multiple UI prototypes in Figma and simplified the interface based on user feedback
Developed the visual feedback system

Technical Implementation:
- Iterated on the prompt for better headline de-biasing

Challenges I Faced

Figuring out Chrome's API
- Challenge: Being an experimental API, there is not too much out there to guide development
- Solution: Used Google Gemini pro to take current documentation and ask it questions
- Learning: Get an LLM to generate examples which you can run and iterate on quickly (chrome dev tools makes this easy)
Language Model Limitations
- Challenge: The API sometimes produced inconsistent results
- Solution: Write a prompt testing suite and collect eval data
- Learning: Build eval suites to A/B test out different prompts and choose the better one
Interface Complexity
- Challenge: Early versions had too many options for headline style and was confusing
- Solution: Stripped away everything but the essential one-click interface
- Learning: Every added feature has a hidden cost in user cognitive load

Technical Implementation

Neutrafy showcases Chrome's AI capabilities through:

Use of the Prompt API for headline generation
Architecture that improves with Chrome's AI advancements

Scalability and Future Potential

Technical Scalability

Zero external API dependencies ensures unlimited scalability
Fully local processing eliminates server costs and privacy concerns
Chrome's evolving AI capabilities will automatically improve performance

Regional Adaptability

Currently supports a few languages to a lesser extent (dependent on Chrome's AI)
Designed for easy language addition as Chrome's AI models improve
Cultural nuances considered in neutralization

Impact and Purpose

Addressing a Critical Need

The project tackles a fundamental challenge in modern society - the increasing polarization driven by emotionally charged news. By providing users with neutral alternatives, Neutrafy helps: