Research & engineering

Hi, I’m Wyett.

I’m a master’s student in Data Science at Harvard, working on reinforcement learning, language models, and the systems that make them more efficient.

Most recently, I worked on LLM training and infrastructure at ByteDance / TikTok.

Wyett Zeng
Wyett “Huaye” Zeng
01

Experience

ByteDance / TikTok

Jun – Aug 2026

Research Scientist Intern · San Jose, USA

Worked on machine learning systems for efficient large language model training.

GPTZero

Sep 2024 – Aug 2025

Machine Learning Engineer Intern · Toronto, Canada

Led development of the transformer-based AI Reviewer and its evaluation and data pipelines. Achieved 88% accuracy within ±1.0 IELTS band and 97% within ±1.5 bands. The work led to a pending U.S. patent on human-aligned LLM review.

Boosted.ai

Jan – Apr 2024

Quantitative Developer Intern · Toronto, Canada

Rewrote the core factor model algorithm, reducing scheduled inference time by 90% and saving 500 compute hours weekly. Updated daily economic factors for more than 20,000 public securities.

CIBC · Gallant MacDonald

Jan – Apr 2023

Data Scientist Intern · Toronto, Canada

Built financial data infrastructure, market analysis tools, and forecasting models for a multi-family office.

Siemens Healthineers

Jan – Apr 2022

Android Software Engineer Intern · Ottawa, Canada

Developed and maintained Android software for the epoc Blood Analysis System.

Doodle Mobile

May – Aug 2021

Product Manager Intern · Beijing, China

Worked on product development and releases for Crossword Quest and Wordsearch.

02

Publications

Google Scholar ↗

ACL 2025 First author

AceCoder: Acing Coder RL via Automated Test-Case Synthesis

Automatically synthesized test cases for reward modeling and reinforcement learning that improve code generation.

Preprint 2026 First author

From Words to Embeddings: Text Representation and the Information Content of Financial News

Comparing keywords, topic models, and LLM embeddings across 135K+ financial news articles to measure how text representation changes the market-relevant information we recover.

03

Projects

Open source

verl-tinker

A Tinker-compatible API server built on verl. Run fine-tuning, reinforcement learning, and distillation on your own GPU infrastructure.

Apartment search

HopIn

An apartment search platform that helps renters find verified listings, compare their options, and decide where to live.

Visit HopIn
04

Education

Harvard University

2025 – 2026 (expected)

Master of Science in Data Science

Expected December 2026 · GPA 3.96 / 4.0

University of Waterloo

2020 – 2025

Bachelor of Computer Science · Artificial Intelligence

Dean’s Honours List every term

Wilfrid Laurier University

2020 – 2025

Bachelor of Business Administration · Finance

Alumni Gold Medalist · Double degree with Waterloo