Last updated
Experience & Education
-
Founding Member of Technical Staff
Vmax AI · San Francisco, CA
I am the first technical hire at Vmax, where we are building automation for RL, starting from RL environments.
- I developed the core training system end-to-end: configuration system that makes experiments agent legible and reproducible, RL infrastructure orchestrating hundreds of GPUs, self-improving software-engineering task generators, and a rollout framework running agents in Harbor tasks on thousands of concurrent sandboxes.
- The system improved an open-source SWE-bench Pro base model by +50% relative under a constrained context window using synthetic tasks.
- I am a contributor to slime and Harbor, covering RL training logic, agent integrations, and sandbox runtime fixes.
-
Research Engineer, MRS
Meta · Sunnyvale, CA
Worked on MRS (Meta Recommendation System), including LLMs for ads recommendation and large-scale user-ad interaction modeling.
- Built billion-parameter autoregressive models for user-ad interaction, improving normalized entropy by 0.07% and earning an org-wide EngEx shoutout as a 0-to-1 use of a new internal ML framework.
- Built an evaluation pipeline for LLM-based CTR prediction, reducing iteration time from hours to minutes and improving AUC from 0.51 to 0.61.
- Ranked as the #1 code contributor in an org of around 40 engineers as a new hire, then was promoted within 9 months.
-
Autopilot Software Engineering Intern
Tesla · Palo Alto, CA
Worked on Autopilot systems infrastructure across build tooling, model-system interfaces, and C++ libraries.
- Built scripts to refactor internal dependencies and reduce related CI time by around 50%.
- Built a high-performance C API for IPC that decouples ML models from Autopilot systems, plus supporting C++ reflection-library work.
- Contributed to Tesla Autopilot's open-source fixed-containers library.
-
M.S. Computer Science
Georgia Institute of Technology · Atlanta, GA
4.0 GPA with coursework across advanced ML, generative modeling, reinforcement learning, algorithms, and systems.
- Teaching assistant for Blockchain and Cryptocurrencies and Machine Learning.
- Coursework included Deep Reinforcement Learning, Deep Learning for Text, Advanced Algorithms and Uncertainty, Efficient ML, Programming Language Design, Brain-Inspired ML, and Information Security.
-
Research Engineer
BTQ · Taipei, Taiwan
Worked on post-quantum cryptography, signature aggregation, zkSNARK recursion, and blockchain protocol security.
- Identified a critical vulnerability in a signature aggregation protocol and prevented a signature-forgery path.
- Built a Rust signature aggregation protocol with around 400x verification speedup.
- Implemented Falcon verifier logic as templated arithmetic graphs and recursively verified it with Plonky2.
-
Quantitative Research Intern
Kronos Research · Taipei, Taiwan
Selected with the best intern assessment score from hundreds of applicants and worked on market microstructure research.
- Built order-tracking tooling to study execution slippage.
- Developed an alpha signal that improved R2 by 42.4% when combined with existing signals.
-
R&D Intern
Sudo Research Labs · Taipei, Taiwan
Worked on blockchain protocol research across DeFi risk, consensus, L2, and cross-chain systems.
- Reviewed DeFi and protocol risks, helping prevent millions of dollars of potential investment losses.
- Contributed open-source documentation to Arbitrum, Polygon Bor, and decentralized-thoughts.
-
B.S. Computer Science and Applied Mathematics
National Yang Ming Chiao Tung University · Hsinchu, Taiwan
Studied computer science and applied mathematics at a top Taiwan CS program, with Presidential Award top-5% performance.
- A+ performance in theory-heavy courses including Advanced Algorithms, Advanced Linear Algebra, Quantum Information and Computation, and Theoretical Aspects of Modern Cryptography.
-
Research Intern
Institute of Information Science, Academia Sinica · Taipei, Taiwan
Worked on graph neural networks and cross-attention models for natural-language inference.
- Represented dependency graphs as matching structure and implemented AllenNLP-compatible model components.
