Imran Nazir

I studied Civil Engineering at BIT Sindri, Dhanbad — graduating with an 8.99 CGPA in 2023. While most of my classmates went into construction, I was in my dorm room learning to code. No CS degree, no bootcamp — just curiosity and an obsession with building things that work.

I taught myself software engineering, data engineering, and then ML infrastructure from the ground up. That journey took me from a Full-Stack intern to a Data Engineer at Episource (acquired by Optum), and eventually to building production ML systems at Optum (UnitedHealth Group) for 3+ years — where I also managed networking for both dev and prod Kubernetes clusters, including firewall security, VPC configuration, and compliance policies.

Today I'm an AI Applied Engineer II at QuillBot, where I own the LLM serving infrastructure for 50M+ users. I work across the full inference stack — from model optimization (LoRA, QLoRA, quantization) to serving layers (vLLM, SGLang, Triton) to orchestration on Kubernetes. I shipped Rengoku, a production Kubernetes sidecar that handles async LLM inference at 5,000+ pages per batch.

I also write. My articles cover GPU inference internals, connection-level overload protection on Kubernetes, and capacity-aware LLM serving. I believe in sharing what I learn in the open.