RSK linear operators and the Vershik–Kerov–Logan–Shepp curve
TO APPEAR Electronic Journal of Combinatorics
Curriculum Vitae
Theoretical machine learning & reinforcement learning · University of Illinois Urbana-Champaign
Research Interests
Theoretical ML, reinforcement learning, online learning, MDP function approximation, sample complexity.
Education
University of Illinois Urbana-Champaign
Aug 2023 – May 2027
Champaign, Illinois · GPA 3.99 / 4.00
- B.S. in Computer Science
- B.S. in Liberal Arts & Sciences, Mathematics (Data Optimization concentration)
- B.S. in Liberal Arts & Sciences, Statistics
Papers
EveGuard: Defeating Vibration-based Side-Channel Eavesdropping with Audio Adversarial Perturbations
PUBLISHED IEEE Symposium on Security and Privacy 2025
Workshop Papers, Manuscripts & Notes
Frozen Policy Iteration for MDPs with Stochastic Transitions under Linear Qπ Realizability
IN PREP Manuscript in preparation
Predictability in Major League Sports: Betting Odds versus Mathematical Models
SUBMITTED Mathematics and Sports (MAS) Journal
Revitalizing Local Democracy: A Human-Centered Audit of LLMs in City Council Journalism
WORKSHOP HEAL Workshop at CHI 2026
Research Experience
Northwestern University
May 2026 – present
Research Intern · advised by Prof. Zhaoran Wang
- Developed a closed-loop LLM meta-orchestrator agent for aircraft wing inverse design, where the agent diagnoses its own surrogate failures across iterative versions and emits corrective source code for the next optimization layer.
- Identified and systematically addressed seven structural failure modes of surrogate-based optimization — including surrogate exploitation, Gaussian trust collapse, family-specific exploitation, and single-objective specification gaming — building a unified taxonomy and corresponding defenses.
- Applied reinforcement-learning-style reasoning to the outer loop: the LLM agent classifies failure modes and upgrades its decision layer, achieving relative robustness gains validated against public RANS datasets (AirfRANS, SuperWing, ONERA CRM).
iDEA-iSAIL Lab
Feb 2026 – present
Research Intern · advised by Prof. Hanghang Tong
- Conducted research in theoretical reinforcement learning, studying sample and computation efficient algorithms for online learning in finite-horizon Markov Decision Processes under function approximation and stochastic dynamics.
- Used concentrability to develop the first sample and computation efficient algorithm for online reinforcement learning in stochastic-transition MDPs under the linear Qπ realizability assumption, resolving an open problem left by prior work restricted to deterministic dynamics.
- Established a PAC sample complexity guarantee of Õ(d2H7C*/ε3) by combining techniques from offline and online reinforcement learning theory, dynamic programming, concentration analysis, and linear function approximation.
- Manuscript in preparation: Frozen Policy Iteration for MDPs with Stochastic Transitions under Linear Qπ Realizability.
Illinois Combinatorics Lab for Undergraduate Experience (ICLUE)
Aug 2024 – present
Research Intern · advised by Prof. Alexander Yong
- Proved a new asymptotic result that the probability that the Schensted insertion algorithm for a uniformly random permutation in Sn exhibits a special "bumping" interaction converges to 1 as n approaches infinity.
- Derived the result using rigorous probability bounds on the RSK linear operator by leveraging the Vershik-Kerov-Logan-Shepp limit shape curve as the central analytic tool.
- Developed proofs spanning real analysis (uniform convergence, limit arguments), probability theory (concentration inequalities, convergence in probability), and asymptotic analysis (bounding combinatorial quantities in the large-n regime).
- Paper to appear in the Electronic Journal of Combinatorics.
Illinois Mathematics Lab
Jan 2024 – May 2026
Research Intern · advised by Prof. AJ Hildebrand
- Researched the data-science perspective of sports matches, analyzing the predictability of outcomes by comparing prediction accuracies across polls, the betting market, Elo ratings, and mathematical models like Bradley-Terry.
- Presented findings at the Rose-Hulman Undergraduate Math Conference 2024, UIUC Undergraduate Research Symposium 2024, and Joint Mathematics Meetings 2025.
- Paper in submission to the Mathematics and Sports (MAS) Journal.
Carnegie Mellon University, HCII
May 2025 – Feb 2026
Research Intern · advised by Prof. Jeff Bigham
- Designed and evaluated an automated journalism pipeline using SOTA LLMs to replicate end-to-end editorial workflows, from transcript segmentation to headline generation and topic prioritization.
- Orchestrated a large-scale crowd-sourced study and demonstrated that LLM-based headline quality and topic prioritization can meet and exceed professional standards.
- Paper in the CHI 2026 HEAL workshop.
National University of Singapore, SERIUS REU
May 2024 – Aug 2024
Research Intern · advised by Prof. Kelvin Fong Xuanyao
- Designed the GARDEN (Generalized Anomaly Recognition and Detection for Enhanced Nurturing) pipeline for early disease detection in farm plants, combining segmentation, generative, and classification models end-to-end.
- Built a leaf segmentation model to isolate regions of interest from complex backgrounds, enabling downstream anomaly detection to focus on plant tissue only.
- Implemented a conditional GAN (cGAN) to synthesize realistic diseased and healthy leaf images, augmenting training data and improving classifier generalization.
- Achieved robust disease classification by ensembling deep learning and traditional ML models, exploiting complementary strengths to improve detection reliability on real farm images.
University of California San Diego
Jun 2023 – Dec 2024
Research Intern · advised by Prof. Xinyu Zhang
- Developed EveGuard, a software-based defense framework to protect voice privacy from vibrometry-based side channels, using a perturbation generator model (PGM) to suppress sensor-based eavesdropping while preserving high audio quality.
- Implemented Eve-GAN, a novel domain-translation task for inferring eavesdropped signals, enabling end-to-end training of the PGM and using few-shot learning to reduce data-collection overhead.
- Achieved a protection rate of over 97% against audio classifiers, hindering eavesdropped reconstruction.
- Paper published in the IEEE Symposium on Security and Privacy 2025.
Grants & Awards
- 2025CMU HCII REU — NSF 2349558
- 2025ICLUE RTG — NSF 1937241
Relevant Coursework
CS: Statistical Reinforcement Learning, Machine Learning, NLP, Deep Learning, Algorithms, Numerical Methods, Numerical Analysis
Math: Honors Real Analysis, Graph Theory, Combinatorics, Abstract Linear Algebra, Abstract Algebra I–II, Algebraic Combinatorics, Linear Programming
Stats: Stochastic Processes, Applied Bayesian Analysis, Survival Analysis
Languages
English: Native
Mandarin Chinese: Fluent, HSK6