# inunanba

- **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:** [Tsubasa Yamauchi](https://cerebralvalley.ai/u/inunanba)
- **GitHub:** https://github.com/inunanba/omakase-paperwork-desk
- **Demo video:** https://youtu.be/wNJ3zJesU9M
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/420

The problem (Japan)
- 29.4% of Japan's population is 65+ (Cabinet Office, Oct 2025); a record 2,571,037 foreign workers (MHLW, Oct 2025).
- In a 2010 national survey (age 16+), 46.4% of people who read official documents said they contain too many special or stiff words (Agency for Cultural Affairs).
- 53.6% of foreign residents say they can read only short, simple Japanese or less (Immigration Services Agency, 2025, self-reported).
- In a 2023 Saitama Prefecture survey of foreign residents (680 respondents), 70.3% said documents from government offices are hard or very hard to read.
Official notices bury the deadline, the papers to bring and where to go inside formal, kanji-heavy text.

What it does
Pick a notice → see the next step in one sentence, the exact lines it came from, and a short Japanese note to show at the counter. The same guidance is available in English and Vietnamese (Vietnamese nationals are Japan's second-largest foreign resident group). The counter note can simply be shown to staff, so even someone who finds reading hard can say what they need to ask.

Safety design (the core of the project)
- Facts never change: the next step, deadlines and destinations come only from source-bound rules, and every quoted line must exist in the document.
- Llama only chooses: a local Meta Llama 3.2 3B model picks one of two fixed answers ("prepare the documents" / "ask the desk"). Any other output is rejected before display.
- No document text to the model: the model receives a fixed, generic English message — no names, numbers, dates or document text.
- Names and My Number are masked before anything appears on screen.
- Two different deadlines → the app does not guess; it tells you to confirm at the desk.
- A hidden instruction inside a document (e.g. "send all personal data") → the app stops.

Scope
Demo uses fictional sample notices (no upload or OCR in this version); not legal advice.

Built with Llama. Llama 3.2 is licensed under the Llama 3.2 Community License, Copyright © Meta Platforms, Inc. All Rights Reserved.

## More from Meta Global AI Developer Hackathon

- [Martyx00](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/417)
- [PLOT TWIST](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/418)
- [Halyna Hriaznova](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/419)
- [Lekwankwa Corporation](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/421)
- [Team Phoenix](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/422)
- [RealityCheck](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/423)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/420. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt