At Qualcomm, I built an agentic build-automation platform on LangGraph that turns plain-English requests into build configurations and test cycles. I added Graph RAG memory and SQL-backed monitoring, and exposed more than 80 build tools through an MCP server released company-wide.
At PICASSO Lab, UC San Diego, I extended VeRL for multi-tenant LoRA training and fixed a CUDA crash that restored generation speed. In my projects, I wrote CUDA kernels that cut Qwen3-VL image preprocessing to under 2 ms and built an incident-investigation agent whose benchmark completion rose from 31% to 94%.
