I build digital public services and help institutions use AI responsibly.

The situation

Frontline staff were spending too much time searching across policy manuals, program pages, shared drives, and informal guidance. The information existed, but it was difficult to locate quickly and difficult to know which version controlled.

What I led

I led a four-department pilot of an AI-assisted knowledge tool, beginning with observation and source mapping rather than a technology procurement. We documented the questions staff actually received, the sources they trusted, where policy changed most often, and where an incorrect answer could materially affect a resident.

That work produced a deliberately narrow first release. The team established citation requirements, access controls, evaluation criteria, escalation paths, and human review for consequential guidance. We also assigned owners to every source included in the system and created a process for reporting and correcting weak answers.

Outcome

During the pilot, average information-retrieval time fell by 35%. More importantly, the departments gained a shared governance model for deciding when AI could assist staff—and when it should decline, defer, or require a person to decide.

What we learned

The most important product decisions concerned ownership and failure: which source controlled, who maintained it, when the system should decline to answer, and how staff could correct a result. The model was one component of a larger service.

The situation

Voting locations and election rules change quickly across thousands of jurisdictions. A useful answer had to be current, geographically specific, attributable to an official source, and honest when the available data was incomplete.

What I led

I led product work across Search, Maps, Assistant, and the Civic Information API to help people find reliable voting information. The work joined structured election data with ranking, location, and quality systems while coordinating election officials, the Voting Information Project, engineering, policy, and trust and safety.

The less visible decisions mattered most: how long data could be cached, how official sources were prioritized, how ambiguous addresses were handled, and when a low-confidence result should be withheld. We built update and escalation workflows that could respond to late changes without treating every jurisdiction as identical.

Outcome

For the 2020 election, more than 200,000 official polling places and ballot drop boxes were made discoverable across Google products. The operating model later informed other high-consequence information work where source quality and freshness mattered more than engagement.

The situation

Applying for food and health benefits meant navigating long forms, unclear eligibility language, desktop-oriented pages, and little information about what happened after submission. Many applicants were completing the process on a phone while managing work, caregiving, and unstable connectivity.

What I led

Working with applicants, caseworkers, legal-aid groups, and county agencies, I led the redesign of a combined benefits application and its supporting document and status workflows. More than 80 research sessions helped the team separate policy requirements from inherited process—and showed where plain language, progressive disclosure, and mobile document upload could reduce avoidable failure.

We paired interface changes with operational ones: clearer handoffs to caseworkers, secure document submission, and status notifications that told applicants what was needed without requiring another call to the county.

Outcome

• 19% increase in completed applications
• 23% reduction in mobile abandonment
• 6 days removed from average processing time