# Sourcenote

- **Event:** [Meta Global AI Developer Hackathon](https://cerebralvalley.ai/e/meta-ai-developer-hackathon)
- **When:** Oct 26, 12:00 PM – Nov 4, 9:00 PM (EST)
- **Where:** Online
- **Team:** [jinsu park](https://cerebralvalley.ai/u/parkjinsu)
- **GitHub:** https://github.com/1jsjs/sourcenote
- **Demo video:** https://youtu.be/_HBXvpVbhfg
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1906

Most AI note apps write the notes for you. That is the wrong job. In a randomized classroom study of 405 students (Cambridge University Press & Assessment and Microsoft Research, 2025), students who took their own notes understood and remembered more than students who only used an AI chatbot. And AI summaries fail quietly: in a 2026 Georgetown and University of Washington study, summaries left out about half of the central details, and people who read a misleading summary answered correctly 44.8% of the time, versus 83.6% after an accurate one.

Sourcenote keeps the student writing and makes the AI the checker. It is a desktop app (Mac and Windows) that turns a lecture recording and its slides into evidence, then checks the student's own notes against it:

- Lecture in, evidence out. The recording becomes time-stamped transcript segments; every slide is read as an image, so a chart counts as evidence, not just its axis labels.
- Bring your own notes. Type or paste them, or import a photo or PDF of handwritten notes. The lines are copied word for word, mistakes included, so the check sees what the student actually wrote.
- What's wrong. Each sentence is matched to its evidence (a moment in the recording, a slide page) and judged against only that evidence. Unsupported sentences are marked "Not in sources" with the reason. In the demo, "the loss falls fastest with a learning rate of 0.001" is caught because the slide 3 graph shows the opposite.
- What you missed. The notes are compared with the lecture's verified key points, and the ones left out are listed with links to the exact moment and slide.
- Cited summary and questions. Every summary sentence cites the recording or a slide and is verified the same way. Questions get cited answers, or "not covered in this lecture".
- 25 languages. Each note has its own lecture language and note language: an English lecture can become Vietnamese notes, and the same mistake is caught from the same slide.

It is built for the students with the most to gain: those studying in a second language, and those who receive note-taking accommodations.

Everything was built during the hackathon: the repository was created on October 27, 2026. Built-in sample notes open without an API key ("Open sample note"). Download: https://github.com/1jsjs/sourcenote/releases/tag/v0.1.0 (the Mac build is signed and notarized; the Windows build comes from CI).

## More from Meta Global AI Developer Hackathon

- [MetaAgents](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1903)
- [Mushtastic Intelligence](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1904)
- [Reflex](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1905)
- [Hog rider](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1907)
- [Murugan](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1908)
- [PollPilled](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1909)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1906. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt