Research ML systems and distributed infrastructure for large-scale model training and serving. My work focuses on efficient LLM-internal observability, resilient collective communication under NIC/link failures, and practical systems that run across real GPU, network, and serving-stack constraints.

I build end-to-end prototypes across CUDA, C++, Python, PyTorch, vLLM, and NCCL-like communication stacks; design GPU-cluster and LLM-serving evaluations; and collaborate on applied ML projects including Internet incident investigation agents and multi-cohort biomedical modeling.

Worked on provenance-based intrusion detection for advanced persistent threats, with a focus on concept drift and adapting to evolving attack behaviors.

I built experiment pipelines, trained and evaluated Transformer-based models on provenance data, and contributed to a cross-institution academic/industry research project that was accepted to NINeS'26.