—Career
Educational Content Pipeline
Structured 20,000+ questions and linked them to exam data problem-by-problem
01Overview
After an early single-giant-prompt design failed, I split the work into three layers by responsibility: Extract → Structure → Meta-tag. A strategy pattern makes each layer independently swappable, so a math PoC expanded to every subject with no re-architecture. I designed and implemented a copyright-aware masking subsystem in collaboration with legal, and built the evaluation on both quantitative metrics and a second-pass check where another AI re-solves each question. It is the flagship of my AI-driven development process (spec → discuss with Claude Code → issues → TDD → iterate with Codex + Claude Code review).
02Highlights
- Structured 20,000+ questions and linked them to exam data problem-by-problem for the first time
- Three-layer pipeline (Extract → Structure → Meta-tag) with independently swappable stages
- Designed and implemented a copyright-aware masking subsystem with legal
- Built an evaluation system where a second LLM re-solves each question to verify output
03Tech Stack
04Related Projects
Shinkenzemi √Route
Founded a 0-to-1 online learning platform for high schoolers, growing from a solo prototype to PM & Tech Lead of a 26-person team.
Challenge AI Learning Coach
Solo-built the Java backend of Benesse's "Challenge AI Learning Coach" — one of Japan's first production EdTech generative-AI products, launched nationwide.