Production AI engineering, opened up to the detail that matters.

This is a lab, not a CV. Each exhibit is a system or tool built to survive production, opened up to the one detail only someone who actually built it would know. The code that backs it is open source.

exhibits

MCP / agent infra

Hardening MCP servers for production

MCP demos work on day one. They break on day 30. These are the field notes from running them for real.

/mcp →

AI security

Hardening a pipeline against prompt injection

Content you did not write can hijack your model. Here is the layered sanitization I run, and exactly what it does not catch.

/injection-defense →

open source

mcp-safe-fetch

A small, single-purpose MCP server: fetch a URL, return injection-sanitized, size-capped text. The reference implementation of the field notes.

/safe-fetch →

currently building

2026-08Shipped a production multilingual dubbing pipeline: speech-to-text → LLM translation → cloned-voice TTS → timeline assembly. Write-up in preparation.
2026-07Document AI on real-world scans: OCR → structured extraction → generated reports, with a vision-model fallback where tables break.
2026-06Open-sourcing mcp-safe-fetch, an injection-aware fetch tool for agents.
2026-06Writing up the MCP hardening field notes as a public reference.
ongoingMore exhibits as each is cleared for public release.