Zhiheng (Carl) Fan

University of Michigan School of Information

Turned failures in AI web automation into product requirements for source checks, confirmation points and error recovery. Co-authored the CHI 2025 workshop paper below.

University of Michigan

Taught SQL, relational design, Python and systematic debugging to more than 300 students over two semesters.

Client-sponsored product project with Quicken

Ran five user interviews and six usability tests, then used them with competitive and heuristic analysis to set priorities for the dashboard, onboarding and tax module.

Team research, University of Michigan School of Information

Designed and built the extraction and provenance workflow of a Chrome extension that links each value an AI extracts back to its source passage; iterated the prompts, schemas and interface on observed failures. In a 12-person comparison with OttoGrid, participants rated its results as more trustworthy (p = .027).

Personal project

Built a searchable map that connects characters from the Touhou Project games to real places and grades the evidence for each connection: a versioned Python data pipeline, a Next.js, React and MapLibre interface on Cloudflare Workers and D1, stable place IDs with version history, and release checks.

Open source

Two fixes to the Windows version, merged upstream: the screen is no longer disabled twice on shutdown (#10197), and a thread releases the thread mutex before it tests for cancellation (#10200).

Open source

Two fixes merged in 0.12.0, each with tests: no dangling commas in creator names (#532), and failed retrievals of recent items reported as MCP errors rather than successful results (#533).

Graduate consulting for Michigan Medicine

Prepared planning inputs with a Python ETL workflow over intervention, staffing and acuity records, and built a linear optimization model to compare staff allocations before an ICU expansion.