Kikigaki Archiver

Inspiration

Local history is often preserved in ordinary memories: the sound of a workplace, the path to a festival, the feel of winter well water, or a recipe learned by watching a relative. These details can disappear before they reach younger generations.

Kikigaki Archiver grew from a real portrait-exhibition practice in rural Japan and asks: how can AI help an interviewer listen carefully and turn one conversation into useful educational and exhibition material without flattening the speaker's voice?

What it does

Kikigaki Archiver guides a Japanese oral-history interview at a calm pace. It asks one respectful question at a time, acknowledges the latest answer, and follows the speaker's concrete details.

After the interview, it generates:

1. an edited transcript that preserves the speaker's phrasing;
2. an evidence-linked life timeline;
3. three exhibition caption proposals; and
4. a downloadable four-page A5 booklet PDF.

The reviewer path uses a fully fictional interview and works without an API key. Live mode uses GPT-5.6 for both interview progression and structured archive generation.

How we built it

The application is a mobile-first Next.js 16 and React 19 web app. Both AI routes use the Vercel AI SDK with openai.responses("gpt-5.6"), explicitly selecting the OpenAI Responses API.

The interview route streams one concise follow-up at a time. The archive route uses a Zod schema to produce structured, validated results. The prompts prohibit inventing dates, relationships, or events, and timeline entries include supporting evidence from the transcript.

The Web Speech API provides optional voice input and read-aloud. React PDF produces the A5 booklet locally with an embedded Japanese font. The fictional demo is deterministic, contains no real participant data, and requires no API key.

OpenAI Codex was the implementation partner. It scaffolded the repository, drafted and refined the prompts, implemented the responsive UI and API routes, wrote tests and documentation, created the demo workflow, and verified desktop, mobile, production, and PDF behavior.

Human decisions determined the Education category, respectful interviewing policy, privacy boundaries, uncertainty handling, and exhibition-oriented outputs.

Challenges

The main challenge was preventing polished output from becoming invented history. Uncertainty is preserved instead of silently filled in, and timeline records expose their supporting evidence.

A second challenge was making the complete workflow immediately reviewable without exposing personal information or requiring an API key.

Accomplishments

• Complete conversation-to-booklet workflow
• Substantive GPT-5.6 use in two core AI stages
• Evidence-linked timelines
• Three exhibition caption proposals
• Printable four-page Japanese A5 booklet
• API-key-free fictional reviewer path
• Mobile, desktop, automated-test, and PDF verification
• No real participant data in the repository

What we learned

Oral-history AI needs a different success criterion from a general assistant. The best next question is the smallest respectful question that helps the speaker continue in their own words. Structured output is also more trustworthy when uncertainty is allowed and supporting evidence remains visible.

What's next

The next phase is a consent-first pilot with exhibition participants, followed by human fact-checking and editorial approval. Future work includes participant-controlled redaction, encrypted storage, deletion controls, archival metadata export, multilingual interviews, and a facilitator dashboard for schools, libraries, and local museums.

Submission baseline commit: 0890138 (July 18, 2026).

Built With

• ai-elements
• codex
• gpt-5.6
• next.js
• noto-sans-cjk-jp
• openai
• openai-responses-api
• react
• react-pdf
• tailwind-css
• typescript
• vercel
• vercel-ai-sdk
• web-speech-api
• zod