So I built Kobunsho Search — an AI-powered search engine that lets you ask questions in plain language (English or Japanese) and get sourced answers directly from these historical documents.

The Problem

NDL's digital collection is a treasure trove, but the search experience is stuck in the early 2000s. You need to know exact keywords in Japanese. OCR quality varies wildly. There's no way to ask "What did GHQ say about land reform in 1946?" and get a meaningful answer.

Researchers, students, and history enthusiasts hit the same wall: the data is there, but it's nearly impossible to find what you need.

How It Works

The architecture is simple:

1. User asks a question in natural language (English or Japanese)
2. Claude extracts search keywords — even English questions get translated to Japanese keywords, since the source documents are in Japanese
3. NDL Labs API returns matching digitized books with full-text OCR data
4. Claude reads the actual pages and generates a sourced answer with citations (document PID + page number)

The entire flow streams the response in real-time, so you see the answer being written as Claude analyzes the documents.

Tech Stack

• Next.js — frontend and API routes
• Claude API — keyword extraction + document analysis with streaming
• NDL Labs API — book search and full-text OCR retrieval
• Supabase — auth and usage tracking
• Vercel — deployment

The Interesting Parts

Bridging the Language Gap

One thing I'm proud of: you can ask questions in English and get answers from Japanese historical documents. Claude handles the translation seamlessly — extracting Japanese keywords for search, reading Japanese OCR text, and responding in your preferred language.

Dealing with OCR Noise

Historical documents + OCR = messy text. Old Japanese typography, damaged pages, and unusual layouts all create noise. The prompt explicitly tells Claude to consider OCR misreadings and interpret from context. This works surprisingly well — LLMs are naturally good at fuzzy text interpretation.

Keeping Costs Down

Each query hits the Claude API twice (keyword extraction + answer generation) and the NDL API multiple times. To keep it sustainable as a solo dev:

• Only the first 5 pages of each matching book are fetched
• Rate limiting per user via Supabase
• Streaming responses so users aren't waiting for a complete generation

Try It

👉 kobunsho-search.vercel.app

Ask anything about Japanese history — try "What were the food rationing policies in postwar Japan?" or "明治時代の鉄道建設について教えて".

What's Next

• Adding more document collections beyond books
• Improving citation formatting with direct page links
• Exploring local embedding search for faster results

Built with Next.js, Claude API, and NDL Labs API. The source documents are provided by the National Diet Library of Japan.

I'm a solo developer focused on niche apps — small tools for specific audiences that big companies will never bother building. Fishing logs, hiking utilities, specialized tools for professionals who are still stuck with spreadsheets.

My stack is Unity and TypeScript, and I use AI-driven development (Claude Code, LLMs) to ship faster than ever.

This blog is about the process:

• Finding niches that are small enough to own but big enough to pay the bills
• Shipping fast as a solo developer with AI tools
• Monetization strategies that work at small scale — no VC, no ads, no growth hacking
• Technical deep dives into Unity, TypeScript, and mobile development

Why Niche?

The math is simple. A mass-market app needs millions of downloads to matter. A niche app needs 1,000 paying users at \(5/month — that's \)60K/year from a single app.

The competition is lower. The users are more loyal. And you can actually talk to your customers.

What to Expect

I'll be sharing:

1. Case studies of niche apps I build (what worked, what flopped)
2. AI workflow breakdowns — how I use Claude Code and LLMs in my daily development
3. Market research methods for finding underserved niches
4. Technical tutorials in Unity and TypeScript

If you're a solo developer tired of competing with Big Tech, follow along. Let's build small and profit.