The name
Kubernetes comes from the Greek κυβερνήτης — the helmsman, the steersman, the pilot at the wheel of a ship. Most Kubernetes content treats that etymology as a curiosity. kubernai takes it seriously: the four surfaces map to a compass rose, the chart-paper aesthetic is the brand, and every article is a navigation chart for one specific operational waterway.
What kubernai is
A reference for platform engineers and ML-platform engineers who already know what a Pod is and are now running (or about to run) AI workloads at production scale. Each article explains one mechanism — how the scheduler scores GPU nodes, how vLLM batches requests, how NCCL detects topology — well enough that the reader can defend a decision about it the next day.
What kubernai is not
Not a tutorial site. Not a vendor-comparison site. Not a news site. Not opinion. Not "5 reasons Kubernetes is dying." Every article slots into exactly one of four surfaces:
- cluster (N) — scheduler, GPU partitioning, autoscaling, node lifecycle
- serve (E) — inference servers, KV-cache, batching, streaming, routing
- train (S) — distributed training, NCCL, checkpointing, fault tolerance
- operate (W) — observability, security, GitOps, cost
How articles get written
Articles are drafted autonomously by a local Qwen 122B model, then scored by a second pass against a reviewer rubric — structural fit, source specificity, API plausibility, decision-frame anchoring, anti-patterns. Drafts that clear the bar publish; drafts that don't queue for a human edit. The bar weights one axis double: did the article invent a CRD field, a metric name, or an API surface that doesn't exist? That's the brand-killing failure mode for a technical corpus, so it's the one the reviewer is least forgiving about.
The pipeline is shared across the kubernai operator's sister sites (visualgrok and pennywhys) — each domain has its own editorial spine; the spine is what the model reads as its system prompt.
Editorial spine, summarized
Every article follows a five-part structure: hook, what-it-is, mechanism, failure modes, decision frame. Every article names at least three real components or projects, includes at least one real YAML / command / metric value, and ends on a specific decision frame — not a generic aphorism. Articles are 900–1,600 words. Outside that range is wrong.
The compass rose
N points to cluster. E points to serve. S points to train. W points to operate. Click any cardinal on the home page rose to navigate to that surface.