The first is our inference control plane — open-weight models and custom task-model zoos, hosted and operated across managed GPU clouds and customer-managed Kubernetes clusters, with scale-to-zero economics, cold-start discipline, and per-token cost accounting that stays correct even when a client disconnects mid-stream — along with the Kubernetes layer those workloads live on: operators, autoscaling, node lifecycle. Familiarity with isolation technology (Firecracker, Kata, gVisor, or comparable), secrets management and egress control (Vault/KMS-class), and hosting stateful systems (vector stores like pgvector/Qdrant, graph stores like Neo4j) with backup and failover discipline.