Power Ruler
2026A power data field for Garmin Edge 1050. Current, average and maximum power sit alongside the power zone, with LOW / HIGH showing which half of the zone you’re in.
I build backend systems and ship reliable software. I also write short essays about growth, responsibility, and the strange business of becoming an adult. Currently open to new roles.
I want to understand how older adults make sense of everyday technology and learn to use it on their own. Research CV (PDF)
I'm a student at UIUC studying Computer Science + Education. Most of my work lives at the intersection of software, learning, and health — I like building things I couldn't find anywhere else.
I grew up planning every detail of my life and slowly learned that you will never feel like you're ready — you can only start being ready. Most of the best things I've done came from starting before I felt prepared: shipping an iOS app, leading a cross-functional AI team, building a wearable for my brother.
Away from the keyboard I climb at Urbana Boulders and Boneyard, cycle on a GIANT Propel, and write short essays about responsibility, readiness, and the gifts of the era we live in. If any of that resonates, say hello.
Built regression tests for a crowd simulator and cut each round from about 40 minutes to 5.5. Used SHA-256 to verify that all 57,754 trajectory rows matched byte for byte across runs.
Maintain the website and respond to inquiries.
Contributed annotations to the Mathematical Derivation Graphs Dataset and developed the MIDGARD Voting Ensemble, combining repeated LLM runs and predictions from different extractors with configurable voting thresholds. Paper submitted to ACL ARR (May 2026).
Working on Social-Emotional-Behavioral (SEB) skills research. Contributing software tooling and data infrastructure that supports the group's studies.
Coursework across systems (CS 225, CS 340, CS 374) and learning science. Interested in the seam between software engineering and how people actually learn.
I'm not a nerd! I have hobbies away from my computer. Bouldering is one of them. I'm an active climber at Urbana Boulders and Boneyard Boulders. If you are interested in Climbing, find me on Instagram!
I also love cycling and have a GIANT Propel ADV SL. If you love ride too, find me on Strava!
A power data field for Garmin Edge 1050. Current, average and maximum power sit alongside the power zone, with LOW / HIGH showing which half of the zone you’re in.
A heart rate data field for Garmin Edge 1050. Current, average and maximum heart rate sit alongside the heart rate zone, with a marker showing your position within it.
Maintainer of a bouldering web platform providing comprehensive climbing beta, route information, and community-driven content for climbers.
A bilingual (English / 中文) curriculum that teaches children critical thinking — motivated by the rise of AI. When answers are cheap and instant, the valuable skill is no longer getting an answer but judging answers and reasoning independently. Two tracks, 24 lessons, plus an assessment, a facilitator hub, and a self-paced course site.
Qwen-VL estimates what's on the plate and how much it weighs; USDA FoodData Central supplies the per-100g macros. Multiply, sum, report. Every provider speaks the OpenAI chat-completions format, so switching is a config change, not a code change.
A local MCP server that gives Claude access to sources public search indexes cover badly: Chinese-language platforms, vertical forums, and community discussion. Seventeen adapters across five clusters, with the verification workflow measured rather than assumed — and the tool descriptions themselves put under measurement by running real models against them.
A personal health knowledge base. Not intuition, not influencers, not "I heard that" — a decision system traceable back to the primary literature. Every conclusion carries an evidence grade and a source (PMID / DOI); every line of every protocol links back to the evidence file it rests on. A claim with no source does not go in.
iOS app built in Swift that integrates HealthKit to transform wearable workout data into readiness and training-load insights, using an on-device, privacy-first architecture.
End-to-end AI web app for automated meeting transcription and summarization. Led a cross-functional team; integrated WhisperX, Ollama LLM, Streamlit UI, and SQLite. 64+ pytest tests at 100% pass rate.
Personal AI agent that fetches news and YouTube video summaries and delivers them via Telegram on a cron schedule. Runs a local Ollama LLM for privacy; supports natural-language config through the bot.
Wearable seizure-prediction device built for my brother, combining advanced sensors and ML to analyze real-time biometric data and deliver timely alerts.
Contributed annotations to the Mathematical Derivation Graphs Dataset and developed the MIDGARD Voting Ensemble, combining repeated LLM runs and predictions from different extractors with configurable voting thresholds. Paper submitted to ACL ARR (May 2026).
A late-night conversation with my brother about whether his daughter — my niece Zeaman — still needs to go to school in an era where the credential economy is quietly breaking down.
Being irresponsible is easy, but it hurts people's feelings.
You will never feel like you are ready, but you can always START to ready.
Two families, one era, and the difference between consuming opportunity and letting it cultivate you.