Experience

Software Engineering Intern (AI Systems)

EUV Tech · Martinez, CA

• Built internal tooling around a production LLM assistant: a data-ingestion pipeline that prepares field diagnostic data for the assistant to import and verify, and an evaluation harness for comparing prompt and model changes reliably.
• Designed a natural-language data analysis feature that turns plain-English questions into sandboxed, editable Python scripts, and hardened it through adversarial testing.
• Moved event aggregation server-side for accuracy, and migrated a knowledge base to a graph database to support path-tracing and centrality queries.

Undergraduate Researcher

University of California, Santa Cruz · Santa Cruz, CA

• Designed and ran experiments on failure modes of vision-language-action (VLA) models controlling robot arms (Franka Panda in robosuite / LIBERO), in collaboration with UC Berkeley, looking at biases induced by model priors and how to mitigate them.
• Tokenized the MS COCO caption corpus and ran a statistical analysis over a 1M-token sample to test whether image tokens follow the same distribution laws as language.
• Built a benchmark of four multi-agent coordination games (Hanabi, Overcooked-AI, Collab Capture, Collab Escape) for evaluating LLM agents; co-authored LLM-Coordination (NAACL 2025 Findings).

LLM-Coordination →

Research Intern

Carnegie Mellon University · Pittsburgh, PA

• Evaluated LLMs as personal assistants on the Enron Email Dataset (500,000+ messages), curating and labeling a subset for evaluation and robustness testing.
• Identified adversarial security risks in language-agent pipelines and proposed mitigations.

Research Intern

University of California, Berkeley · Berkeley, CA

• Studied grounded language learning for efficient video understanding, implementing multimodal architectures in PyTorch.
• Built a pipeline that finds the video frames answering a user's natural-language query, and presented the work three times, including as a poster at the SACNAS national conference.

Project page →