Agentic internal tools

The tools that shape the work

Good design systems don't run themselves. Behind every component library and documentation site is a layer of tooling that makes the whole thing usable. For much of my time at Meta, building and refining that layer was my job.

I joined Meta's internal tools team embedded within an organization that supported design systems across the company. Handoff was inconsistent, documentation was hard to maintain, and as my role shifted to the Facebook design system, AI capabilities were entering the product roadmap faster than teams could make sense of them. The tools I built shifted from solving workflow friction to changing how designers and engineers collaborate entirely.

Owning the documentation hub

I owned the internal documentation site that served as the single source of truth for every design system at Meta, covering design guidance, component specs, accessibility standards, and framework-specific availability across the company. From running vision sprints with leads representing Meta's top design systems to shipping over a dozen features, I maintained and evolved a product used by thousands of internal users daily.

Ensuring content was machine-readable was also part of this work, making detailed design system guidance accessible not just to designers browsing the site, but to the AI agents being built on top of it.

Design System Agent

I collaborated with a small team to launch an internal AI agent trained on the Facebook design system, including the documentation site I had owned. Mapping the flow of a design system request from submission through evaluation, I identified where human judgment was essential and helped build a structure that could hold up as usage scaled.

I coordinated multiple dogfooding sessions to train the agent. We scored its responses to real questions and used those findings to build a repeatable framework for assessing AI-generated output. The goal was to make AI-generated answers something that could be trusted, not just tried.

Shipping a SQL dashboard in hours

When the Quality Month program I organized needed a real-time participation dashboard, I built it from scratch in a few hours using natural language prompts. The work spanned the full stack: SQL queries, GraphQL, React, responsive CSS, accessibility, and dark mode. The dashboard replaced manual tracking entirely and made an interactive leaderboard available to hundreds of internal participants competing for the best score. Thousands of UI, accessibility, and content issues were logged and fixed during Quality Month.