cv

B.S.

University of California, Los Angeles (UCLA)

Computer Science

• Data Structures & Algorithms
• Neural Networks and Deep Learning
• Deep Learning for Computer Vision
• Foundations of Computer Vision
• Probabilistic Models in Genomics
• Signal Processing
• Introduction to Robotics
• Advanced Neural Networks and Deep Learning (Graduate)
• Computational Imaging (Graduate)
• Human Factors in AI (Graduate)

Embodied Web Agents: Bridging Physical-Digital Realms for Integrated Agent Intelligence

NeurIPS 2025 Datasets and Benchmarks Track (Spotlight)

Co-authored a benchmark and dataset for evaluating agents that integrate web interaction with embodied/physical context, supporting research on grounded, multimodal, and agentic intelligence.

Multimodal Safety Research Intern

A10 Networks

Conducted research on multimodal safety alignment for vision–language systems, focusing on failure cases in visual grounding, efficiency constraints, and adversarial robustness. Designed and evaluated a vision–language guardrail model and initiated a scalable multimodal safety dataset derived from large pretraining corpora.

• Characterized failure patterns across grounding, latency, and coverage dimensions.
• Implemented evaluation harness for robustness under visually distracting inputs.
• Established end-to-end data pipeline toward public multimodal safety release.

Undergraduate Researcher

Chang's Group at UCLA NLP

Studied long-context memory in embodied multi-agent systems, developing Temporary-LoRA–based mechanisms to improve knowledge persistence in VirtualHome simulations. Led large-scale synthetic dataset generation for multimodal agent training and co-authored a benchmark on embodied web agents.

• Temporary-LoRA adaptation for context retention in multi-agent settings.
• Generated 10k+ first-person, multi-step embodied trajectories.
• Co-author, NeurIPS 2025 Datasets & Benchmarks Spotlight.

Undergraduate Researcher

UCLA StarAI Lab

Worked on structured reasoning for generative models, including constraint-guided decoding and discrete optimization. Reproduced structured diffusion results, built SDD-based decoding constraints for language models, and analyzed unbiased gradient estimators for Mixture-of-Experts routing.

• Reduced decoding-time constraint complexity via SDD compilation.
• Empirical comparison of implicit vs explicit structure in generation.
• Variance analysis of multi-sample estimators for discrete expert selection.

Software Engineer Intern

CreatorHub

Developed full-stack MVP features for a creator–brand platform, spanning backend APIs, database design, and authentication workflows. Emphasized schema correctness, security, and production-ready deployment.

• Delivered user-facing product features under startup timelines.
• Defined MongoDB schemas with validation for consistency guarantees.
• Implemented secure credential handling and recovery flows.

Robotic Manipulation with Robosuite

Trained and evaluated reinforcement-learning agents for simulated robotic manipulation using Robosuite, comparing PPO and SAC under sparse and dense reward regimes and diagnosing instability sources in exploration.

• PPO: ~90% task success after 10M environment steps.
• SAC: full task success in <1M steps.
• Identified reward-sparsity–induced mode collapse failure mode.

Hallucination Detection

Proposed a hallucination-detection framework for medical question answering with retrieval-augmented LLMs, integrating model uncertainty and retrieval agreement to support safer EHR-based reasoning.

• Designed confidence score combining entropy, factuality, and retrieval similarity.
• Formalized hallucination categories for clinical QA evaluation.
• Benchmarking protocol for trust calibration in medical RAG systems.

Novel View Synthesis

Analyzed classical and neural view-synthesis methods, comparing light-field rendering, NeRF variants, and 3D Gaussian Splatting with respect to efficiency, memory footprint, and visual fidelity.

• Comparative analysis across NeRF, Mip-NeRF 360, and splatting methods.
• Identified regimes where Gaussian Splatting dominates in efficiency.

Webmaster

Bruin Spacecraft Group

Led a team of 6 to redesign and maintain the club’s website, improving usability for sponsors and prospective members. Coordinated with 7 admin teams to ensure accurate project documentation and progress updates.

Instructor and Mentor

exploretech.la

Designed and taught workshops on computer vision, generative AI, and NLP for 100+ students at exploretech.la’s annual outreach expo. Mentored high-school students from Title I schools in Los Angeles, introducing core ML concepts such as linear models and CNNs.