At Microsoft, I develop agent workflows that turn execution traces into structured memories and reusable skills. This work improved relevant-memory retrieval by 18% in benchmarked repeat-task evaluations.
I’ve also optimized semantic retrieval and LLM inference, reducing median retrieval latency by 20% and inference costs by 12%. I integrated Microsoft Agent Framework with Azure AI Foundry models for multi-step workflows involving task decomposition, memory retrieval, tool invocation, and outcome verification.
At NVIDIA, I contributed to a personalized recommendation platform for retail applications. I prepared interaction and product data, created behavioral features, and assisted with hybrid recommendation models and embedding-based candidate retrieval.
I supported GPU inference optimization using CUDA, TensorRT, and Triton Inference Server, contributing to approximately 30% lower prediction latency. In personal projects, I’ve developed agent-based code review workflows and a risk and demand forecasting system.
