I build auditable AI workflows and privacy-first developer tools. My current focus is turning document-heavy or API-heavy manual processes into small, reviewable systems with explicit validation, visible failures, and human approval.
- text-layer PDF, text, JSON, and CSV extraction with source evidence; Excel workbooks are review outputs rather than input files, while scanned PDFs and OCR are separately scoped;
- Python and API workflow automation with tests, logs, and handoff notes;
- local-first AI or retrieval prototypes where sensitive decisions remain human-owned;
- workflow audits that separate deterministic checks from model-assisted steps.
Typical pilot outcomes are evidence.xlsx plus review_queue.csv for extraction, or changes.xlsx plus an approval-ready summary.md for old/new document comparison.
The first two seven-day pilots are $149 / ¥999: one representative data source, up to 20 documents or 500 rows, one bounded workflow, one structured output, one human approval point, one correction round, deployment notes, and seven days of defect support. I use only public, properly redacted, or synthetic samples during the pilot.
Purchase the seven-day pilot · View the auditable workflow demo · Open a public scope inquiry
- regulated-workflow-demo — local-first evidence extraction, controlled-document diffing, human review queues, audit logs, synthetic fixtures, and passing CI.
- quietloop — evidence-driven, privacy-first reflective automation for personal agents.
- LinkDock — a native macOS companion for Android screen mirroring and remote control over trusted local connections.
I do not claim that a prototype is production-deployed, legally compliant, or guaranteed accurate without evidence. Legal, medical, financial, compliance, and other consequential decisions stay with the responsible human reviewer.