Skills and projects

On-device dictation app for macOS

Role

Product owner and sole developer, with AI coding agents

Team

Solo

When

Sep 2026 – present

Hold a key, speak, and cleaned-up text lands at the cursor in any app. Speech recognition and cleanup run fully on device, in Chinese and English.

• Text appears a median 1.4 s after the key is released (172 dictations, 6–8 November 2026); 2,400+ dictations of my own daily use so far
• Memory down from 1.7 GB to 96 MB
• Learned vocabulary raised correct terms on a finance-call test set from 907 to 974 of 1,193
• 230+ merged pull requests and 40 decision records in 7 weeks, with a regression suite and a nightly benchmark on 7 speech test sets
• Figures as of 8 November 2026.

Autonomous household recall agent

Role

Sole builder, with AI coding agents

Team

Solo

When

Aug 2026Google's All Things Agentic Hackathon, Taskmaster track

Four Google ADK agents (Gemini) check three live US federal recall feeds (CPSC, NHTSA, openFDA) every hour. When something a user has registered is recalled, they pursue the remedy claim, with no chat and no user present.

• False alarms treated as the core risk: a rule-based prefilter that cannot flag on its own, then an LLM judge for each candidate pair
• 20 real decoy products give zero false positives
• Running agent calls concurrently cut a sweep from 173 s to 2 s

VR emergency-response training

Role

Product Manager, level designer, API developer

Team

Team of 5

When

Jan–May 2025University of Michigan, EECS 440 capstone

Stakeholder research with the University of Michigan Fire Marshal and an analysis of the 2013 Yarnell Hill Fire located the training gap in crisis communication. That became LLM-powered voice conversation with the characters in the scenario, through the OpenAI API.

• Wrote the MRD, PRD and BRD: 3 personas, 5 competitors, pricing and a 15-month roadmap
• Ran the Jira backlog and sprints
• Shipped on Meta Quest 2 and 3 at 72 fps
• The team's projected 3-year financial model: break-even at 27 months, 105% ROI

University of Michigan course-registration redesign

Role

UX researcher and designer

Team

Team of 4

When

Jan–Apr 2025University of Michigan, EECS 493

4 interviews (80+ data points) showed students juggling 6+ platforms to register for classes; one concept-testing participant described "thirty-something tabs" open at once. The answer is a Figma prototype with 2 end-to-end flows, from finding a course to registering for it.

• 7 think-aloud sessions over two rounds
• 14 usability findings ranked by severity, leading to 12 design changes: for example, flagging a time conflict while a section is being chosen, not after
• All 4 new users in Round 2 completed both core tasks

A campus-navigation needs assessment and a store usability test

Role

Sole researcher

Team

Solo

When

Sep–Dec 2024University of Michigan, SI 422

Two studies I ran alone, from consent form to report. Five recorded interviews with Michigan students, using photo elicitation, showed that the core problem was finding the way indoors: connected buildings, room numbering, entrances. Then three moderated remote usability tests of ktown4u.com, a Korean K-pop store that sells to buyers abroad.

• Interview transcripts coded into an affinity diagram: 148 quotes, 20 codes, 4 groups; delivered a persona and a journey map
• Usability test: a test plan, a 10-task facilitator guide and per-task logging; 40% overall task success across 3 participants
• Version comparison, shipping-cost calculation and order optimization failed for all 3 participants; fixes ranked by impact and effort

Food-recall search over FDA and USDA records

Role

NLP pipeline lead

Team

Team of 5

When

2025University of Michigan, EECS 486

A 7-class taxonomy of recall reasons over 27,500 FDA and USDA records, with named-entity recognition for pathogens.

• BERT classification at 81.57% macro-averaged precision, measured on 900 hand-reviewed records
• A scraper that recovered 1,509 USDA records the official API could not return
• Pathogen extraction with scispaCy

Real-time multiplayer conversation card game

Role

Solo project

Team

Solo

When

2025

A web app (PWA) with two modes for 2–6 players and 71 prompts in 3 depth tiers, kept in sync live on Supabase.