I build data pipelines and AI tooling where governance is part of the architecture, not an afterthought — PII controls that run before persistence, quality checks that actually stop a pipeline, and documentation that states limitations honestly instead of overselling.
| Project | What it is |
|---|---|
| Wikipedia attention archive (Hugging Face dataset) | The 1,000 most-viewed Wikipedia articles per day in 13 languages, from July 2015 to yesterday. A Mac mini at home collects it every morning. It also lists topics that jumped abroad while Japanese Wikipedia stayed quiet. CC0. |
| hf-signal-dashboard · live page | Buy/sell markers for every Nikkei 225 and Nasdaq 100 stock, each printed next to how that marker did in a backtest. Public data only (Yahoo, SEC EDGAR). Not investment advice. |
| taka-pixel-reference · live page | A one-page reference for Pixel phones, in Japanese: update deadlines, known bugs per Android version, recovery steps. |
| edinet-mcp | Retired in July 2026. It served Japanese company financials to MCP clients; a free service (edinetdb.jp) now covers the same ground, so I stopped maintaining it. Left up as a record. |
- Controls as code — a report nobody must act on is not a control; a gate with a non-zero exit code is.
- Honest threat models — pseudonymization is not anonymization, and READMEs should say so.
- Primary sources — numbers come from filings and official data, with attribution, or they are marked as missing.
Python (pandas, pytest) · TypeScript (Node, MCP SDK, Ajv, Vitest) · GitHub Actions · SQLite


