I'm an AI-focused builder in Tokyo, turning ideas into small, useful systems. I enjoy understanding language models from first principles, connecting AI to real tools through MCP, and building thoughtful native apps for macOS.
Current focus: practical AI agents, Model Context Protocol, local-first workflows, and learning by rebuilding the fundamentals.
| Project | What it explores | Built with |
|---|---|---|
| TaskPaper MCP Server | Gives AI assistants structured access to TaskPaper on macOS | TypeScript · MCP |
| LLM in Racket | Rebuilds Karpathy's llm.c approach for minimal, CPU-only GPT training |
Racket · GPT |
| LLM in R | Implements a GPT-2 training framework in pure R | R · GPT-2 |
| PDF Outline Skill | Lets Codex add PDF bookmarks from printed tables of contents | Python · Codex |
AI systems → agents, tool use, context, evaluation
LLM internals → tokenization, training loops, inference
Local-first → private workflows that stay close to your data
Human + AI → tools that amplify judgment instead of replacing it
- 🧠 I learn best by rebuilding things from the ground up.
- 🛠️ I like software that is focused, local-first, and genuinely useful.
- 🌏 Based in Tokyo, working across Chinese, English, and Japanese contexts.
- 📈 My background in data analysis keeps me curious about how models really behave.
Explore my projects · Stack Overflow · LeetCode
Build to understand. Ship to learn.