Deep-dives and talk write-ups from running 150+ Kubernetes clusters, 500+ GPUs, and the platforms on top of them — written down so you don't have to learn it the hard way.
From my Pod Weekly live EKS demo: ResourceGraphDefinitions turn YAML + CEL into a CRD plus micro-controller at runtime, the dependency DAG is inferred rather than declared, and kro + ACK puts cloud and cluster in one reconciliation loop. With deck.
↗ gpu infrastructure · study · 8 min readStop wasting 70–90% of your GPU capacity. Hardware slicing with MIG + DRA, smart queuing with Kueue cohorts, groupless provisioning with Karpenter + KEDA, and a GPU reaper for zombie pods — from static GPUs to an elastic compute fabric.
↗ gpu infrastructure · deep-dive · 10 min readA four-layer reference architecture for LLM inference on Spot: capacity acquisition as market-buying, GPU-aware scheduling vs the fragmentation tax, the workload contract, and utilization as a control loop. With the anti-patterns named.
↗ fleet operations · deep-dive · 11 min readRings bound risk, phases encode dependency order, verification defines done. EKS lifecycle economics, the two-phase waiter every async control plane needs, and the eight-step shape that turns any fleet mutation into a batch job.
↗ gitops · deep-dive · 10 min readFleet GitOps is a data-modeling problem: the (cluster × addon × version) matrix has to live somewhere. ApplicationSets, multi-source composition, controller sharding math, and the failure modes — sync storms, schema debt, the kubectl escape hatch.
↗ ai × infrastructure · deep-dive · 12 min readWhy "chatbot with kubectl" fails, and the subsystems that fix it: ReAct with a structural critique gate, context engineering that cuts tokens 10×, Langfuse-gated evaluation, and a write path that's a security architecture, not a feature flag.
↗ observability · deep-dive · 9 min readOrchestrator state and user experience are different facts. Where you probe from changes what you measure, the prober's failure model, cardinality as the second bill, and wiring probes into SLO burn rates.
Drafts in the queue: Karpenter bin-packing for GPU workloads and ArgoCD notifications at 100+ project scale. Follow on LinkedIn for new posts.