selected as one of twenty founding engineers in pear vc's inaugural pear prime program, out of 500+ applicants. building an ml-powered internal tool that helps pear source and evaluate deals using the firm's proprietary deal-flow data.

research assistant and technical point-of-contact for the airlab vision-language model team, benchmarking model architectures on masking, video understanding, and multimodal reasoning for the darpa triage challenge. building bayesian network models for probabilistic reasoning research, and designing no-code annotation interfaces that let medics give structured feedback to medical ai without writing code.

first author on "3d dataset distillation: condensing 3d datasets for enhanced data efficiency in 3d reconstruction," presented at the evga workshop at isca 2024. designed and built nerd in pytorch — a system that prunes redundant views from nerf training sets using projection-validity and depth-consistency constraints, cutting required training data by up to 60% with near-full reconstruction quality. advised by prof. yingyan (celine) lin.

developed methods for unlearning artist styles from latent diffusion models to address copyright and data attribution in generative ai. co-author on an arxiv paper on diffusion-model unlearning.

evaluated ai-generated datasets for hardware design workflows. built tooling to quantify and compare key hardware design metrics across generated datasets.