Inspiration

I started from the ELI5 idea: turn complex text into something a child can truly understand—often as a story. From there, I generalized it so the same input can make sense to any reader from age 5 to expert without hype or guesswork.

What it does

Elix delivers personalized comprehension on three pillars:

  • Reader-aware: age personas (5/10/15/Adult/Expert) tune tone and vocabulary.
  • Purpose-aware: Story, Summary, TLDR, Key Points.
  • Context-aware: optional background; outputs stay grounded to the provided text.

How we built it

Elix‘s inference is an ordered pipeline built with the Gemini API, utilizing Gemini Nano's specialized Writer and Summarizer models. It starts with optional Web Search for context, then language detection. A dynamic prompt combines the original text, background context, user's reading level, and output style, sent to the appropriate model. It's a responsive web application.

Challenges we ran into

  • Making age tone reliably affect Summary/TLDR/Key Points, not just Story.
  • Reflection loop blew latency/tokens; shipped single-pass; defer deep reflection to Flash.
  • Translation API broke bullets; chunking avoided format loss but caused cross-chunk term drift.

Accomplishments that I'm proud of

I'm proud I successfully built an MVP that demonstrates true "personalized comprehension," effectively adapting content to individual needs beyond basic summarization within the Hackathon timeframe. Successfully implementing the "ordered pipeline" for managing Gemini Nano's specialized models was a significant achievement, resulting in a functional, intuitive user experience.

What I learned

I learned that personalized AI requires precise prompt engineering, balancing style instructions with model creativity. Using Gemini Nano's specialized Writer and Summarizer models for specific tasks proved highly efficient. Structured context integration is also key for model performance.

What's next for Elix

  • Prompt API for user-defined modes.
  • Use Gemini Flash/Pro with reflection to improve output quality or more complex task.
  • History & sharing (exploration) to compare, revisit, and share results.

Built With

Share this project:

Updates