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 |
|---|---|
| edinet-mcp | MCP server for Japanese company financials (FSA EDINET). ~3,600 listed companies, normalized to English keys and JPY. No API key, no signup. Runtime schema validation and data-freshness warnings built in. |
| bpo-contact-center-intelligence-pipeline | GRC-first contact-center ETL (synthetic data). Keyed HMAC pseudonymization and a fail-closed quality gate placed before the warehouse — the ordering is the control. |
- 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