Curriculum Vitae

Carnegie Mellon University

Pittsburgh, PA

• Relevant Coursework:
  • Introduction to Deep Learning (PhD-level)
  • Distributed Systems
  • Generative AI (PhD-level)
  • LLM Systems
  • AI Agents
• Honors & Awards:
  • 3^rd/238 Automatic Speech Recognition
  • 5^th/266 Face Verification

Fu Jen Catholic University

New Taipei City, Taiwan

Smile Line AI

San Francisco, CA

• Scoped the AI engine for Smile Line's flagship Smile Lite MDP app with the founding team, from model and data choices to EU and US privacy rules, ahead of deploying it to the company's device and app
• Benchmarked detection models and quantized the chosen YOLO detector to run chairside on hardware the company already ships, so it checks whether each photo is framed correctly while the clinician shoots

Carnegie Mellon University, Language Technologies Institute

Pittsburgh, PA

• Ported course homeworks to JAX, Flax, and Optax for TPU training with Google, for 300+ students, and authored an LLM pretraining assignment to build a Llama-style model from scratch with RoPE, SwiGLU, and RMSNorm

Carnegie Mellon University

Pittsburgh, PA

• Designed a fail-closed policy engine that keeps a governed LLM agent from answering off-corpus and an evidence gate that withheld 296 of 300 claims under speaker overlap with full coverage on clean audio

Future Educators

Washington, D.C.

• Built an OpenAI-backed RAG advising assistant over 1.2M+ GPT-labeled federal charter school records and deployed it on AWS as a FastAPI service at p95 under 2 s for 300+ daily sessions, 47% faster than manual lookup
• Shipped an LLM document-processing pipeline on Claude 3.5 Sonnet that reached 93% extraction accuracy on a human-labeled eval set and cut manual review time by 70%

The Heroes Garage

Boston, MA

• Reduced bid error against manual appraisal by 37% with a gradient-boosted model over Carfax history and past sale prices that set the buying team's live bids at Manheim and ACV auctions
• Deployed a repair cost model trained on past repair orders that cut dealer estimate error (MAPE) from 22% to 7%

Creator & Maintainer

1k+ Downloads

• Wrote a Rust terminal app where coding agents such as Claude Code and Codex build in parallel and review each other's work

Creator & Maintainer

180+ Stars · 3k+ Downloads

• Brought Jupyter notebooks natively into the terminal and over SSH via a Rust backend on ZeroMQ and a Lua UI

Contributor

6.8k+ Stars

• Fixed FlashInfer's Mamba selective state update CUDA kernels so head dims that are not multiples of 64 compile and run correctly