One GPU, Two LLMs

Building a production-grade LLM inference platform on EKS with GPU time-slicing, KServe, Knative, and a custom gateway. Finally, debugging everything that broke.