At Amazon, I develop recursive Transformer architectures and run end-to-end pre-training, followed by supervised fine-tuning and LoRA adaptation to investigate recursive computation, parameter sharing, and model behavior.
I also develop and curate multilingual training datasets for Alexa. Controlled ablation studies showed improved performance, particularly on challenging evaluation sets, and I developed Sherlock to distinguish model failures from evaluation-pipeline issues and support failure-driven data generation and cleaning.
At Apple, I designed Conduit, a multi-agent LLM framework for interacting with enterprise infrastructure, and built orchestration pipelines integrating Kubernetes, Splunk, and CI/CD systems. My research includes peer-reviewed publications in IEEE and Springer, including work on cricket shot classification and traffic sign detection.
