CV

Master's

Carnegie Mellon University

Master of Intelligent Information Systems (MIIS)

GPA: 4.14/4.33

• Advanced Natural Language Processing (A+)
• Multimodal Machine Learning (A+)
• Prompt Engineering (A+)
• Introduction to Question Answering (A)
• Independent Study in Language Technologies (A)
• Directed Study in Language Technologies (A)

Bachelor's

University of Southern California

Computer Science and Applied & Computational Mathematics

GPA: 3.98/4.00

Minor in Artificial Intelligence Applications. Specializations in Applied Analytics and Video Game Programming (USC Viterbi ITP). Dean's List, USC Dornsife and USC Viterbi (2021-2025).

• Language Models in Natural Language Processing (A)
• Applied Machine Learning for Natural Language Processing (A)
• Applied Neural Networks (A)
• Capstone: Design and Construction of Large Software Systems (A)
• Mathematical Statistics (A)
• Probability Theory (A)
• Numerical Methods (A)

Social Gym and SPaRTan: Benchmarking and Improving LLM Social Reasoning via Multi-Agent Game Tournaments

Preprint

Social Gym is a benchmark of 21 multi-agent social games with verifiable, objective outcomes; SPaRTan is a training-free self-play and reflect-transfer loop that produces transferable playbooks improving LLM social reasoning across games and models.

Believing without Seeing: Quality Scores for Contextualizing Vision-Language Model Explanations

ACL 2026

We introduce Visual Fidelity and Contrastiveness -- two explanation quality scores that help users more appropriately rely on vision-language model predictions without seeing the image.

RECAP: An End-to-End Platform for Capturing, Replaying, and Analyzing AI-Assisted Programming Interactions

ACL 2026 Demo

RECAP captures AI chat and code edits in VS Code, merges them into replayable timelines, and provides analysis modules for studying developer-AI interaction patterns.

ELI-Why: Evaluating the Pedagogical Utility of LLM Explanations

Findings of ACL 2025

Evaluate the pedagogical utility of LLMs in tailoring explanations to users with different educational backgrounds.

Attributing Culture-Conditioned Generations to Pretraining Corpora

ICLR 2025

This paper introduces MEMOed, a framework to analyze whether AI generations are driven by memorization or generalization, with a focus on cultural symbols.

Enhancing Debugging Skills of LLMs with Prompt Engineering

Technical Report

Study on improving LLM debugging through prompt engineering techniques. Evaluated few-shot learning and chain-of-thought approaches on GPT-3.5, revealing limitations in current debugging capabilities.

Social RLVR

Carnegie Mellon University; ongoing follow-up to Social Gym with Prof. Maarten Sap

• Developing multi-agent social scenarios with rule-computed outcomes for reinforcement learning with verifiable rewards (RLVR).

Social Gym and SPaRTan

Carnegie Mellon University; with Prof. Maarten Sap

• Built Social Gym, a Python, game-agnostic N-agent environment with finite-state game logic, partial observability, and rule-defined rewards across 21 games, enabling reproducible multi-turn evaluation without LLM judges.
• Ran role-balanced tournaments for seven closed- and open-weight LLMs across 17 competitive and four cooperative games, covering 560+ episodes per model; fit Bradley–Terry/Elo ratings and role-conditioned capability profiles.
• Designed SPaRTan, a training-free play–reflect–transfer loop and evaluated within-game iteration, cross-game and multigame transfer, and distillation into six student models.

RECAP

Carnegie Mellon University; with Prof. Tongshuang Wu

• Built an open-source VS Code instrumentation and replay platform in TypeScript that captures Copilot chats and fine-grained shadow-git edits, links AI suggestions to code, and supports extensible behavior analysis.
• Deployed RECAP with a cloud data pipeline in a Spring 2026 CMU software engineering for machine learning course, capturing 2,034 prompts and 8,239 code edits from 41 students across 406 work sessions.

Attributing Culture-Conditioned Generations to Pretraining Corpora

USC INK Lab; with Prof. Xiang Ren

• Contributed large-scale corpus analysis, Python data engineering, and visualization to MEMOed, a framework for attributing culture-conditioned generations across 110 cultures.
• Built Google APIs survey infrastructure and a routing service that directed participants to culture-specific forms for scalable data collection.

ELI-Why

USC INK and DILL Labs; with Profs. Xiang Ren and Swabha Swayamdipta

• Co-created a 13.4K-question benchmark for testing whether LLM explanations adapt to elementary, high-school, and graduate-level learners; led task, rubric, and prompt design and built the human-study web application.
• Ran two human studies across 811 participants and multi-metric analyses, showing GPT-4 explanations matched their requested grade level 50% of the time versus 79% for human explanations.

Believing without Seeing

USC DILL Lab; with Profs. Swabha Swayamdipta and Xiang Ren

• Designed Visual Fidelity and Contrastiveness, two quality scores for VLM explanations, and evaluated their calibration against model correctness on A-OKVQA, VizWiz, and MMMU-Pro.
• Designed and ran Prolific studies with 300 annotations per dataset and condition; the combined scores improved users' correctness judgments by 11.1 percentage points and reduced false belief in incorrect predictions by 15.4 points.

Machine Learning Engineer Intern

Adobe

• Shipped 5+ agent skills via an agent skill generator merged into the team's shared skill repo, turning rough developer requests into standards-compliant skills grounded in internal prior art and self-audited against the agentskills.io spec.
• Developed the systematic quality measure for the team's agent recommendation pipeline, a reference-free 6-dimension rubric combining LLM-as-judge with deterministic checks, reproducible to 0.01–0.05 per-dimension variance across repeated runs and surfacing two systematic weaknesses for the team to act on.
• Took that measure to production across the org: ported it onto the internal GenAI evaluation service as a reusable scorer and staged it end-to-end for two enterprise customers, on a four-stage Databricks pipeline built to run it.

Data Engineer Intern

CMind

• Built a LangChain-based RAG prototype over an internal analytics DB (chunking, embeddings, FAISS indexing/retrieval, LLM answer synthesis) with metadata filters and notebook/API utilities.
• Designed CEO-speech trait metrics and an auto-evaluation pipeline with custom rubrics, achieving Fleiss' κ of 0.89–0.91 against human annotations.

Graduate Teaching Assistant, 11-711 Advanced Natural Language Processing

Carnegie Mellon University, Language Technologies Institute

• Supporting a graduate NLP research course covering modeling and learning algorithms and culminating in an original research project.

Teaching & Grading Assistant

University of Southern California

• Coordinated logistics and grading for two computer science courses (CSCI-102 and CSCI-360), serving approximately 300 students per term.
• Led weekly office hours and discussions for approximately 20 students, clarifying concepts and providing feedback on assignments.

VLM Hallucination Mitigation

Training-free latent reflection for vision-language models, at Carnegie Mellon University.

• Built a training-free latent reflection module that monitors decoding entropy and injects reflection steps when the model is uncertain.
• Benchmarked against DAMO, VCD, OPERA, and DeGF on POPE/RePOPE/MMBench across LLaVA-1.5-7B, InstructBLIP-7B, and Qwen2-VL-7B.
• On RePOPE, raised LLaVA-1.5-7B accuracy from 84.0% to 90.3% with the lowest calibration error (ECE 0.028) among all methods compared.

LLM Prompt Recovery

Developed a system to recover user prompts using fine-tuned Mixtral models, achieving top 3.4% in a Kaggle competition.

• Achieved a score of 0.65 using sentence-T5-base and sharpened cosine similarity.
• Ranked 75/2175 globally in the Kaggle competition.
• Published fine-tuned model on Kaggle for broader access.

AI-Based Career Advisor

Built an AI tool to assist users in planning career paths based on skills and interests.

• Developed a Streamlit-based interactive UI integrating GPT-4o for job suggestions.
• Implemented cosine similarity search for skill-job matching.
• Integrated Bing AI for real-time job application link retrieval.

Enhancing Debugging Skills of LLMs with Prompt Engineering

Improved LLM debugging capabilities through advanced prompt engineering techniques.

• Experimented with various prompting strategies to enhance debugging efficiency.
• Achieved significant improvements in LLM performance on debugging tasks.

Automated Hate Speech Detection in Social Media

Developed an advanced ML model for detecting hate speech, achieving 94% accuracy.

• Fine-tuned BERT for classification tasks.
• Enhanced online safety and inclusivity through robust model optimization.

USC CURVE Research Fellowship

USC Viterbi School of Engineering

Awarded in four terms (S24, Su24, F24, S25); total funding $6,750.

Silver Medal, Kaggle Competition

Ranked 75/2175 (Top 3.4%) on the global leaderboard, LLM Prompt Recovery Project.

USC Academic Achievement Award

Awarded in Fall 2022, Spring 2023, Spring 2024, and Fall 2024. Covered 11 units of tuition costs in total, amounting to approximately $24,000.

4th Place, USC Integral Bee Competition

Ranked 4th in the USC Integral Bee Competition.

1st Prize, International Linguistics Olympiad

Senior Level, Individual Open Round, China.

1st Prize, International Linguistics Olympiad

Senior Level, Team Open Round, China.