Posts on platform decisions, GPU scheduling and LLM infrastructure, written from work I have done. Also available as an RSS feed.
A KServe predictor sat Pending for three hours requesting a GPU that nothing in Git asked for. The injector was a Kubernetes defaulting rule most people have never read - and GitOps was structurally unable to show it.
A thought leadership piece on why AI cost attribution needs to happen at the protocol level, not as an afterthought. References MCP, Kagenti, and the emerging need for agent-aware FinOps.
An opinionated take on migrating a TB-scale observability platform from the ELK stack to AWS Managed OpenSearch for 200+ microservices at PVH Europe, and when managed beats self-managed.
An opinionated take on building production infrastructure for LLM-powered healthcare products — why I moved from a dedicated cluster to a shared one, CloudFront VPC Origins over API Gateway, and KEDA over HPA for async LLM workloads.
A technical deep dive into designing Kortex, a multi-provider inference gateway with intelligent routing, circuit breakers, OpenTelemetry tracing, and cost tracking.