Case study · 2026

AI workflows for support teams

Make information available to an agent without losing the responsibility to review it.

← All selected work

The work

At Automattic, I built agents for triage, research, drafting, and self-review. I authored and collaborated on read-only MCP tools for billing, ticket tags, and attachments, including image compression for screenshots.

A boundary that mattered

I wrote a shared library that removes secrets and personal data before model calls. When the checks can’t run, it fails closed. Tools on two other teams adopted it.

From problem to shipped change

I also traced WooPayments Analytics converting an already converted amount a second time, and shipped a merged fix. That investigation sits alongside the tools: both begin with a problem someone needs resolved.

A high-level account of employer work. The contribution includes collaboration; internal tickets, interfaces, and customer information aren’t reproduced here.

Read the merged WooPayments fix

Screens and interaction states

Triage → Research → Draft → Review

An explanatory sequence of the functions above. This is not an internal product screen or an exact production architecture.