Proof of work
Shipped, not just slideware
Selected projects, with my actual role and the hard part of each.
ASUS Clinical AI Assistant · 智慧病歷助手
Live in production
@ VGHTPE · since 2025-11-12 · co-developer
ProblemClinicians spend hours daily on documentation — note-writing, summarization, discharge prep — work that demands judgment but rewards almost none.
My roleClinical workflow bridge: shadowed residents to surface real needs, worked with ASUS engineering on LLM prompting + EMR integration, aligned attendings on acceptance criteria. The cross-team alignment was the hardest and most decisive piece.
EngineeringHelped make a constrained model usable in real clinical workflows: mapped key clinical data fields and workflow requirements, reduced unnecessary context burden, and worked with ASUS engineering to shape a dependable documentation pipeline — turning "not usable" into "usable in real workflows."
OutcomeIn active clinical use at VGHTPE since Nov 2025. Publicly launched at the 2025 Taiwan Healthcare Expo (Dec 4–7, 2025).
LLM prompting
EMR integration
Stakeholder alignment
Real hospital deploy
Public coverage: ASUS × VGHTPE →
CLR-voyance — Clinical Reasoning Evaluation
arXiv · under submission
2026 · co-author (4th of 8) · arXiv:2605.09584 · under submission (NeurIPS 2026)
WhatReinforcing open-ended clinical reasoning for inpatient decision support with outcome-aware rubrics. The 8B model scores 84.91% on the CLR-POMDP benchmark — ahead of GPT-5 (77.83%) and MedGemma-27B (66.66%).
My roleClinician-side validation: contributed to the clinical-alignment study — helping curate per-case rubrics, grade candidate responses, and give blinded pairwise preferences. The human-judgment layer that decides whether a model's reasoning is actually safe and sound.
Why it mattersThe research side of the same thesis: clinical AI must survive clinician judgment, not just benchmark scores.
Clinical reasoning eval
Rubric grading
LLM
Read on arXiv →
CT Annotation MVP
Open source
2026 · clinical AI · solo build
What117 normal anatomical structures auto-segmented on abdominal CT. Runs locally on Apple Silicon — no cloud upload, which reduces PHI exposure risk.
ValidationInformally reviewed by a radiologist and a medical physicist at VGHTPE; in their initial assessment, several structures were comparable to local commercial tools.
TotalSegmentator v2PyTorch MPS
117 organs
local · PHI-minimizing
View live demo →
Presentation Coach
Private preview
2026 · LLM agent + voice TTS · co-built with Matt
WhatA voice-aware coach for academic talks: analyses pace, audience attention, and structure, then returns evidence-grounded recommendations — with exemplar audio generated from user-provided reference audio, with consent.
Hard partNot the LLM — the voice. Ran 7 head-to-head TTS routes against Whisper-verified ground truth; landed on ElevenLabs Instant Voice Clone (6.5s of reference → a transcript-faithful clone, even cross-lingual). Tested on 3MT-style talks, including examples from UCL, SUNY Buffalo, and U Tokyo.
Claude agentElevenLabs IVCvoice clone
Also built: Mitral Valve Dynamics — a 10-module interactive simulator (React + Vite); and Miracle Shiba — an NFT project founded in a single sophomore winter, where a generative algorithm composed 10,000+ unique pieces. Range from research-grade clinical tooling to zero-to-one product.
