Junior at The Harker School focused on machine learning, competitive programming, and mathematics. Independent ML researcher applying machine learning to real-world problems, with professional experience in investment research. Works in Python, C++, Java, Docker, Git, etc.
EDUCATION
The Harker SchoolSeptember 2024 – June 2028
Freshman GPA 4.23 (weighted)
Sophomore GPA 4.43 (weighted)
Cumulative GPA 4.33 (weighted)
Advanced Placement curriculum:
AP Calculus BC (5), AP Physics C: Mechanics (5)2024
AP Computer Science A (5), AP Chinese (5)2025
AP Chemistry (5), AP Microeconomics (5), AP Macroeconomics (5), AP European History (5)2026
AP Biology (..), AP United States History (..), AP Physics C: E & M (..)2027
COSMOS @ UCLA — Cluster 6: From Medicine to Hollywood: AI for Speech and ImageryJuly – August 2026
California State Summer School for Mathematics and Science — selective 4-week residential STEM research program
Machine learning for speech processing and computer vision, with applications spanning medical diagnostics to entertainment
USACO Courses w/ X-Camp (602H)
Topics include dynamic programming, graph and tree algorithms, DSU, MST, and combinatoricsJanuary 2025 – Present
RESEARCH
Bias Transfer in Large Language Models through Supervised Fine-Tuning ➤ October 2025 – Present
Independent Research, sole author — accepted to the AACL-IJCNLP 2026 Student Research Workshop (peer-reviewed, non-archival)
Designed and ran a controlled fine-tuning study on Llama 3.2 3B Instruct across 3 corpora (NELA-GT, NELA-PS, and US high-school Wikipedia articles), producing and evaluating 4 models on 1,200 completions across 15 salient-topic prompts
Scored output bias with an LLM-as-Judge rubric and Word Embedding Association Test (WEAT) effect sizes, isolating overfitting and content hallucination as primary drivers of measured bias
Found that corpus-level bias scores do not predict fine-tuned model bias, showing why LLM-as-Judge evaluations must separate factual accuracy from biased framing
Developing a Machine Learning Algorithm for Wikipedia Vandalism Detection ➤ Summer 2024
Conducted with mentorship from Dr. Françeska Xhakaj of Carnegie Mellon University, through Pioneer Academics
Built a linear-regression classifier on the English-Wikipedia PAN-WVC-10 dataset (Amazon Mechanical Turk labels), achieving 80% accuracy, 76% precision, 89% recall, and a 13% false-positive rate
Demonstrated scalable automated detection of malicious edits for the largest volunteer-edited online encyclopedia
PROFESSIONAL EXPERIENCE
Next Capital — Operations / Investment Intern
➤
July – August 2025
Researched and authored investment reports on target sectors, informing leadership decisions on emerging opportunities
Produced underlying research for 2 promotional videos and 3 articles that expanded Next Capital’s market visibility
Conducted market research and field analysis to support incubated portfolio projects
ACHIEVEMENTS
USACO Gold Division2024 – 2025 Season
Calico Gold Brick (top 11 US, top 16 pre-college, top 20 int'l)2025 Spring
Writing USACO-style competitive programming problems for HPC hosted programming competitions.
GPL 2026, "Concert Planning II"
Boxuan is an aspiring K-Pop star planning her concerts in California. The state can be thought of as a forest of N cities with M (0 ≤ M ≤ N − 1) roads between them. A forest is a collection of one or more disconnected components that are tree graphs.
Each city has fans, who are either excited or tired of her singing.
When Boxuan performs in a city v, word travels and influences only the cities neighboringv.
a city with excited fans attends the concert, and returns home tired;
a city with tired fans is rejuvenated by her melodic voice, and returns home excited.
The residents of city v are not affected, as they are boycotting her concert over noise pollution concerns.
Boxuan knows the binary state of her fans at each city today, and which ones need to be excited / tired by the end of her tour.
A tour is defined as a set of cities, for each of which Boxuan performs a concert exactly once. Tours A and B are distinct iff Boxuan visits a city in tour A but not in tour B, or Boxuan visits a city in tour B but not in tour A. The tour may be empty.
Count the number of tours that can get her from the state now to her goal. This count can be large, so report it modulo 998 244 353, or 0 if no such tour can do so.