
Cloud-Native Inference & GPU Splitting
LLM inference needs CUDA Graph optimization, but CUDA Graphs freeze allocation at capture. TEN’s RA:X team solves this in place with GRIP, accepted to SoCC ‘26.

Nginx Ingress to Cilium: Migration Retrospective
We migrated a production Kubernetes cluster's CNI from flannel to Cilium after Ingress-NGINX reached end-of-life. Here are the five issues that surfaced over ten days after removing kube-proxy — SNI, timeouts, NodePort, monitoring — and the lessons learned.
