CV

Research Intern — RL for LLM reasoning

IBM Research

• Ran ablations and reproductions of existing self-distillation pipelines and uncovered failure modes of OPSD and SDPO.
• Ran a prefix-advantage study across Qwen3-1.7B/4B/8B characterizing where privileged-teacher signal localizes in a rollout. Submitted to the NeurIPS 2026 workshop on Transitioning from Pre-Training to Post-Training.
• Investigating other uses of distillation — whether privileged information can be used in a helpful, non-cheating way.

Researcher — RL for generative models

MIT-IBM Watson AI Lab

• Investigated RL fine-tuning of flow models for molecule and crystal generation.
• Co-author on Tailored primitive initialization is the secret key to reinforcement learning (ACL 2026); built the embedding-similarity pipeline used for the paper’s diversity analysis, showing Tailor’s warm-start data is more reasoning-diverse than the Standard-CoT and 4-STaR baselines.

WITTI Wintern — quantitative finance

Hudson River Trading

Software Engineering Intern — multi-agent systems

Talkdesk

• Prototyped agentic user memory for a multi-agent system and a multi-agent evaluation framework for a production RAG pipeline.

Researcher — AI security & agents

MIT Media Lab

• Built attestation for AI agents negotiating inside GPU-backed trusted execution environments; wrote circom circuits producing zkSNARK proofs that the negotiation ran as specified.
• Built tool-using agents (Gmail, Google Calendar, Shopify APIs) and an evaluation harness for multi-agent negotiation.

Researcher — generative AI for design

MIT DeCoDE Lab

• Built the sketch-to-3D retrieval pipeline (2D sketch → nearest 3D car mesh) and benchmarked retrieval models and similarity metrics; co-author on the resulting ASME IDETC-CIE 2025 paper.

Researcher — network science

MIT PRIMES-USA

• Proposed and evaluated an algorithm combining path-based link prediction with influence-maximization heuristics to identify future vital nodes in weighted collaboration networks.
• Introduced RA-2 and quasi-local RA-2, new link-prediction similarity metrics; built the simulation and evaluation pipeline (7 link-prediction metrics × 12 centrality measures × 8 selection algorithms, under simple and complex contagion); wrote the initial manuscript.
• Co-author (alphabetical order) on papers in Physica A (2026) and on arXiv (2024).

B.S.

Massachusetts Institute of Technology

Computer Science (Course 6-4, Artificial Intelligence and Decision Making)

• GPA 5.0 / 5.0. Expected May 2027.
• ML / AI: 6.7960 Deep Learning, 6.7920 Reinforcement Learning: Foundations and Methods, 6.8300 Advances in Computer Vision, 6.7800 Inference and Information, 6.S184 Generative AI / Diffusion Lab (IAP); in progress: 6.8611 Natural Language Processing, 6.5940 TinyML and Efficient Deep Learning Computing, 6.4210 Robotic Manipulation
• Theory and math: 6.1220 Design and Analysis of Algorithms, 18.404 Theory of Computation, 18.701 Algebra I, 18.702 Algebra II, 18.510 Introduction to Mathematical Logic and Set Theory, 18.03 Differential Equations
• Systems: 6.1800 Computer Systems Engineering, 6.1910 Computation Structures, 6.1903 Introduction to Programming in C and Assembly, 6.1010 Fundamentals of Programming; in progress: 6.4400 Computer Graphics
• Other: 6.4590 Foundations of Information Policy, 24.09 Minds and Machines, 14.01 Principles of Microeconomics, 21H.152 Modern China, 24.02 Moral Problems and the Good Life

Tailored primitive initialization is the secret key to reinforcement learning

Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (ACL 2026), Volume 1, pp. 33300–33318

Influence prediction in collaboration networks: An empirical study on arXiv

Physica A: Statistical Mechanics and its Applications, 689, 131451

AI agents in engineering design: A multi-agent framework for aesthetic and aerodynamic car design

Proceedings of the ASME 2025 IDETC-CIE, Volume 3B, 51st Design Automation Conference, V03BT03A048

The social sphere model: Heuristic influence prediction in evolving networks

arXiv preprint arXiv:2402.03522