Writing

Notes from the infra stack.

Deep dives on GPU orchestration, CI/CD, Kubernetes, observability, and the operational practices behind production AI infrastructure.

  1. 7 min read

    Dynamic PR environments for services on EKS

    Per-PR Kubernetes environments on a shared EKS cluster — ArgoCD git-directory generators, a 3-layer Helm value cascade, multi-source Applications, and a namespace-reaper controller for the cleanup ArgoCD pruning misses.

    • gitops
    • eks
    • argocd
    • kubernetes
    • helm
  2. 1 min read

    vLLM vs SGLang vs TGI — operational benchmarks (coming soon)

    Side-by-side production benchmarks of the three open-source inference servers — throughput, latency, KV-cache behavior, and the operational gotchas you'll hit at scale.

    • inference
    • vllm
    • sglang
    • tgi