Interests

What I Build

ML systems end-to-end: model training, evaluation pipelines, RAG tooling, on-device inference, and the CI/CD infrastructure that gets it all to production reliably. At Apple, I shipped production ML across multiple iOS releases and built Python-based CI/CD systems for dataset validation, model regression testing, and release quality gates.

AI Evaluation + Multimodal Systems

• Multi-turn LLM safety evaluation, co-advised by Prof. Rosanna Bellini and Prof. Damon McCoy : built an automated harness to collect and evaluate chatbot behavior across thousands of multi-turn conversations, using LLM-as-judge scoring validated at Cohen's Kappa 0.80–0.85.
• Multimodal video and data pipelines for robotic policy learning with Prof. Lerrel Pinto, NYU CILVR: 3× dataset generation via diffusion-based augmentation, cross-modal grounding with CLIP and VLMs, imitation learning workflows in JAX.

How I Work

I care about the full path from model behavior to user impact; where latency hides, how failures are caught, what it takes to ship AI outside a notebook.

Education

Outside the Code

Basketball, table tennis, and photography, preferably in a city I've never been to before.