BookMirror Agent

Inspiration

BookMirror was built from a simple problem: reading with AI often becomes a summary, but real learning needs a loop. A reader needs to choose a small passage, hear a focused guide, respond in their own words, and keep a memory of what they learned.

What it does

BookMirror Agent turns a selected book segment into a guided reading session.

The demo shows:

• source-based guide generation with Gemini
• guided discussion that responds to the learner's answer
• Gemini TTS audio generated from the accepted guide transcript
• Chinese and English reading flows
• Notion writeback for learning memory
• a no-preload flow where the guide starts from the selected passage, not a pre-recorded audio clip

The goal is to help the reader think with the text instead of receiving a fixed answer.

How we built it

BookMirror is built with Next.js and TypeScript. Gemini is used for guide generation, discussion, and reflection shaping. Gemini TTS is used to turn the accepted guide transcript into audio. The app keeps the selected source branch visible so Chinese and English sessions stay separated.

The learning record is written back into Notion as structured memory: source context, learning note, and progress state.

Challenges

The hardest part was making the demo honest. I removed the old preloaded audio path and made the flow start from the selected source. I also fixed English audio language locking and made transient Gemini response-shape errors retry-safe instead of showing raw provider warnings.

What I learned

The most important design lesson was that an AI reading product should not replace the reader's thinking. It should narrow the context, ask better questions, and preserve the learner's own interpretation.

Built With

• cloud
• genkit
• next.js
• notion
• react
• typescript
• vercel
• zeabur