Work
I like working on AI systems and the infrastructure around them — retrieval, evaluation, data pipelines, and the parts that connect models to real users.
Each entry below opens if you want the detail.
Education
Fifth-Year Master's, Language Technologies Institute
Expected August 2027Carnegie Mellon University, Pittsburgh, PA
Current coursework Hide coursework
Search Engines (11-642), AI Agents (11-768), Multimodal Machine Learning (11-777), Introduction to Research in Language Technologies (11-705), Directed Research (11-910)
BS in Information Systems and Computer Science
May 2026Carnegie Mellon University, Pittsburgh, PA
Minor in Logic and Computation
GPA: 3.9/4.0 (Major GPA: 3.9/4.0)
Selected coursework Hide coursework
Machine Learning (10-301), Generative AI (10-423), Natural Language Processing (11-411), Algorithm Design and Analysis (15-451), Distributed Systems (15-440), Introduction to Computer Systems (15-213), Parallel and Sequential Algorithms (15-210), Bug Catching: Automated Program Verification (15-414), Foundations of Software Engineering (17-313), Database Design and Development (67-262), Probability and Statistical Inference I–II (36-235/236), Game Theory (80-305)
Experience
Data Engineer Intern
TikTok
Data tooling for internal AI agents — finding the right data source from a plain-language question, and diagnosing why livestream revenue moved.
What I worked on Close
- — Designed and, with my mentor, implemented an enterprise data-discovery capability that helps internal AI agents locate usable data sources from natural-language questions, rather than requiring users to already know table names.
- — Packaged it as a self-contained, agent-callable Python service that returns a stable data-source identifier along with the evidence behind each suggestion, and integrated it into the existing company data agents.
- — Worked on a livestreaming operations revenue-diagnosis workflow: detect revenue downside, diagnose which metrics moved abnormally, interpret the business cause, and suggest corrective action. Handled the data extraction, cleaning, metric aggregation, feature engineering, and scheduled pipeline behind it.
- — Both systems were validated with operations and business stakeholders and adopted into their working process.
Research Assistant
Chimps Lab, CMU
Red-teaming and bias-auditing research, including PersonaTeaming — a CMU–Apple paper at UIST 2026.
What I worked on Close
- — PersonaTeaming (CMU–Apple collaboration, UIST 2026), second author: suggested persona as an additional axis for automated red-teaming and worked on the dynamic persona-generation mechanism, which generates a candidate persona per seed prompt and keeps whichever one better fits the attack.
- — Implemented the end-to-end workflow, set up the experiment orchestration, and ran the evaluation across six target models, measuring both attack success and corpus-level diversity. Wrote the technical results section of the paper.
- — Put together the PersonaTeaming Playground interface with persistent database storage and deployed it on AWS; it was used as the platform for an industry practitioner study run by the Apple side.
- — WeAudit: worked on tooling to analyze demographic distributions in AI-generated images for bias audits, parallelized large-scale image generation with batching for reliability, and containerized and deployed the system on AWS EC2.
Research Assistant — NLP-side lead
Sapling Lab, CMU
How people form and revise beliefs about an AI’s mind over a long conversation.
What I worked on Close
- — NLP-side lead on an NSF-funded interdisciplinary research program studying Mutual Theory of Mind — how the mental-state attributions users form toward socially responsive AI emerge and evolve over sustained interaction.
- — Working on an ongoing first-author simulation study, which grew out of an idea I wanted to pursue, on how user characteristics, assistant social behavior, and conversational situation shape mind attribution across a multi-turn conversation. Manuscript in preparation.
- — Worked out the simulation pipeline and the turn-by-turn measurement approach for tracking attribution trajectories; it is intended both as a paper and as a computational tool supporting the broader research program.
Software Engineer Intern
AlpSoft, Singapore
Secure enterprise web work for a Singapore government agency.
What I worked on Close
- — Maintained and upgraded secure enterprise sites for EnterpriseSG; shipped UI/UX and backend API improvements.
- — Remediated penetration-test findings by removing exposed URL paths and tightening authorization on user and admin actions.
- — Supported deployment sessions on AWS using Docker and Terraform.
Full Stack Engineer & Team Lead
Enterview (Startup), Pittsburgh, PA
An AI mock-interview platform, with a team of four.
What I worked on Close
- — Worked with a team of four on an AI-driven interview-preparation platform with real-time mock interviews.
- — Put together a multi-agent mock interview session and used AI-assisted development to speed up feature delivery.
- — Optimized WebSocket handling for lower latency and smoother feedback.
Research Intern
Tsinghua University, THUNLP — OpenBMB
Organizing heterogeneous REST APIs so a language model can pick the right one to call.
What I worked on Close
- — Prototyped embedding-based organization of heterogeneous REST APIs for the BMTools tool-learning framework, standardizing API documentation into a consistent representation.
- — Clustered tool representations so a language model could select an appropriate category before attempting a call, and wrote unit tests for tool-call format and output readability.
Teaching Assistant
CMU — Great Ideas in Theoretical Computer Science (15-251)
Tutoring proofs and theory, one student at a time.
What I worked on Close
- — Support students through 1:1 tutoring sessions on proofs and theory topics.
- — Design study guides and exam reviews.
IS Consulting Project
ProtectPT (Non-Profit), Pittsburgh, PA
A searchable database for a local anti-fracking non-profit.
What I worked on Close
- — Co-designed and delivered a searchable database supporting anti-fracking advocacy.
Projects
Dealer's Choice
June 2025 – PresentAn online multiplayer poker platform for the variants we play at home games.
Play it: dealer-s-choice.vercel.app ↗
What I worked on Close
- —Custom game variants with real-time multiplayer gameplay.
- —CI/CD pipeline for automated testing and deployment.
- —Deployed on an ECS cluster in private subnets behind a public ALB.
- —Optimized WebSocket connections for low-latency player interaction.
- —The rules for the variants are on my personal page.
Publications & Writing
PersonaTeaming: Persona-Conditioned Automated Red Teaming
UIST 2026 · CMU–Apple collaboration · Second author
Conditioning automated red-teaming on personas, so the attacks reflect the range of people who actually use these systems.
Adds a human-perspective axis to automated red-teaming by conditioning adversarial prompt mutation on personas, including a dynamic method that generates and scores candidate personas rather than relying on a fixed set. I suggested the persona axis, worked on the generation mechanism, implemented the workflow, and ran the evaluation across six target models.
Solver-Guided Fine-Tuning for Accurate and Explainable Poker Decisions
10-423/623 Generative AI, Spring 2026 · with Cassandra Shi & Harry Huang
Fine-tuning an 8B model to explain its poker decisions, not just make them.
A course project fine-tuning Llama 3 8B to jointly produce a poker action and a natural-language explanation, using solver-backed labels from PokerBench as ground truth. Since no human explanations exist for these scenarios, we prompt a larger model as a teacher, score its candidates with a custom rule-based verifier, and validate that verifier against 200 blind human judgments. Supervised fine-tuning closes most of the zero-shot gap (18.0% → 56.2% exact match), and preference optimization (DPO) sharpens explanation quality further at a small cost to action accuracy — an alignment tax we trace back to the action and explanation objectives sharing the same backbone parameters.
Read the paper (PDF)Mutual Theory of Mind in Multi-Turn Human–AI Conversation
First author · Manuscript in preparation
A simulation of how mind attribution shifts, turn by turn, across a conversation.
A simulation study of how user characteristics, assistant social behavior, and conversational situation shape mind attribution turn by turn. Still ongoing at Sapling Lab.
Skills
- Languages
- Python · TypeScript · JavaScript · C · Standard ML · SQL · HTML/CSS
- Backend & Systems
- NestJS · Django · Flask · REST APIs · WebSocket · Distributed systems · Concurrency · Redis
- Cloud & Infrastructure
- AWS (EC2, ECS) · Docker · Terraform · CI/CD · Cloud security
- ML & LLM systems
- PyTorch · Hugging Face · LoRA / PEFT · SFT · DPO · RAG · Embeddings · LLM evaluation · Agent tooling
- Data
- PostgreSQL · MySQL · SQLite · Data pipelines · Feature engineering · Metric aggregation
- Quantitative & Formal
- Probability & statistical inference · Game theory · Algorithm design · Formal logic · Program verification