The form everything else is built on. A one-person studio shipping cloud platforms and AI products, quietly, until they're ready.
Accelerator-aware clusters on EKS and GKE: NVIDIA GPU Operator, MIG and time-slicing, DRA-native device allocation (ResourceClaims, now GA in Kubernetes), Karpenter GPU node pools, and Kueue or KAI-style queueing and gang scheduling, so expensive silicon stays busy.
Production model serving with vLLM, llm-d, NVIDIA Dynamo and KServe: continuous batching, KV-cache-aware routing and offloading, prefill/decode disaggregation, autoscaling on tokens and queue depth, not just CPU.
Curated internal model catalogues across Vertex AI Model Garden, Bedrock and open-weight registries, behind one gateway for routing, fallbacks, guardrails, quotas and per-team cost attribution.
Evaluation harnesses, tool-use and MCP integration, tracing and regression suites that turn agentic workloads into something you can measure, compare and ship with confidence.
Tuned Linux images, drivers and kernel parameters, eBPF-based networking and observability, and hardened node images for GPU and CPU fleets.
Multi-cloud AWS and GCP foundations in Terraform, CI/CD pipelines that make shipping boring, and resource right-sizing and cost engineering for AI-heavy bills.
Infrastructure, observability and a deploy path come before features. Everything after fits better.
Good platforms are felt as speed and calm, never as ceremony.
The AI landscape changes weekly. Decisions rest on this month, not last year.
Schuhwerk is German for footwork, the craft of making shoes. A last is the form a shoe is shaped around: invisible in the finished product, and the reason it fits.
That is the job of a platform. Kubernetes clusters, pipelines, data flows and AI systems that nobody notices when they work. Ghostlast is that craft, offered to teams that need a platform they can trust.