Blog

A practical redesign of a 4,000-GPU B200 cluster using RoCEv2 over Ethernet. This companion dives into lossless fabric engineering, BGP underlay, congestion control, and real-world trade-offs versus InfiniBand at hyperscale.

I built a leaf-spine data center fabric, connected it to Kubernetes with Cilium BGP, exposed services through LoadBalancer VIPs, enforced L3/L7 policy, and traced the end-to-end data path with Hubble, from client request to pod response.

Step-by-step tutorial for running a multicluster Kubernetes setup locally using Kind. Covers regional LoadBalancer routing, cross-cluster connectivity, failover simulation, and rolling updates across three clusters.

Your organisation's knowledge is locked in PDFs and documents nobody has time to read. This guide shows you how to train a private AI on that material, run it on your own machine, and query it like a senior expert who never sleeps.

A practical walkthrough of building a sovereign AI agent stack on Kubernetes using open-source tools. Model inference, memory, tooling, and observability run entirely on-premises, keeping prompts, telemetry, and data inside infrastructure without relying on external AI APIs.

A complete walkthrough of the network fabric design for a 4,000-GPU NVIDIA B200 cluster: 3-tier InfiniBand backend, dedicated Ethernet fabrics for management and storage, OOB network, rail-optimized topology with intra/inter-rail communication patterns, power, cooling, and rack layout.

Cilium VXLAN, FRRouting eBGP, Hubble, and Kubeshark on a 4-node ARM64 cluster. Covers routing limitation, the FRR self-peering problem, and a 52-check validation suite from scratch.

Build a production-ready 4-node Kubernetes cluster on Ubuntu 24.04 LTS — from bare VMs to Cilium CNI with Hubble observability. Includes every gotcha we hit so you don't have to.

Is your AI fabric blind? Static ECMP can’t handle Elephant Flows. Learn how BGP NNHN & Global Load Balancing provide the "X-ray vision" needed to stop congestion and maximize your GPU investment.