At NVIDIA, I defined and implemented a production AI platform for technical-support workflows, combining retrieval, tool calling, structured outputs, and human escalation. In controlled testing, its agent workflow reached 87% successful multi-step task completion.
I designed a hybrid RAG pipeline that improved Recall@5 from 74% to 89%, and built evaluation harnesses for groundedness, task success, latency, and model cost. I also fine-tuned compact instruction models with LoRA adapters, reducing median inference cost by approximately 31% without materially changing held-out task accuracy.
At Meta, I owned ranking model initiatives for content-discovery products. My work included PyTorch ranking models and feature pipelines that improved offline NDCG@10 by approximately 8%, as well as serving safeguards supporting 99.95% service availability.
Earlier, I developed fraud-risk models and production scoring services at Capital One, worked on clinical-operations data and NLP at Pfizer, and built backend integrations and SQL analytics at Verizon. Across these roles, I’ve worked on model evaluation, deployment, monitoring, and responsible operation in regulated and high-volume environments.
