AI is the largest vertical on the blog. It covers what a team needs to take generative models (LLMs) into production on sovereign systems: the complete LLMOps pipeline, the inference engines, the data, the training and the observability.
👉 Start here: Anatomy of an LLM request in production — the full tour of a single real request crossing the six stages of the LLMOps pipeline. If you read only one post in this vertical, read that one. It works as a mental map for everything else.
How the vertical is organised
Posts are grouped by sub-topic (visible through tags):
llmops— the six-stage pipeline, request anatomies, landscape pieces, OSS versus hyperscaler catalogues.llm-inference— KV cache, PagedAttention, disaggregated serving, vLLM on Kubernetes, multi-tenant GPU cluster.llm-data— ingestion with PostgreSQL + Qdrant, RAG over Kafka, data versioning with DVC and lakeFS.llm-training— continuous fine-tuning with LoRA, cyclic retraining, prompt versioning.evalsandsafety— golden sets, LLM judges, jailbreak resistance, PII, prompt injection.llm-observability— OTel tracing, MCP observability, eBPF for drift detection.
Where to begin
If you have never worked with LLMs in production, start with MLOps for LLMs in 2026 and then The six-stage LLMOps pipeline. If you already operate an LLM system, jump straight to the forensic walk through a real request: Anatomy of an LLM request in production.
For the OSS versus cloud decision, the two complementary catalogues: OSS versus hyperscalers (AWS, GCP, Azure) and The OSS catalogue tool by tool.
For platform engineers
Several posts cover the GPU + Kubernetes infrastructure that holds up LLM inference: vLLM on Kubernetes, LLM operators on K8s and Multi-tenant GPU cluster. These live in AI but carry the kubernetes tag, so they also show up when filtering by that tag.