Tous les articles
Articles techniques sur la construction de la plateforme minicloud — GitOps, Kubernetes, IA et ingénierie bare-metal.
How we migrated the minicloud root CA private key from a plaintext Kubernetes secret into HashiCorp Vault's PKI engine — importing the same CA so no certificate had to be re-issued, no trust store had to be updated, and no service saw any disruption.
Most tutorials teach you to deploy on AWS. We went the other way: five bare-metal nodes on the floor, Cloudflare at the edge, and cloud services only where physics makes self-hosting impossible. This is why.
Most enterprises carry 25 years of identity debt — AD on-prem, LDAP connectors, Entra sync agents. We built a Kubernetes platform from scratch and landed directly at the modern end of that arc. This post explains the difference and why it matters.
How to build a production AI gateway that routes between local models and cloud APIs, enforces department-level access controls, strips PII before it leaves your network, and traces every call — all behind a single OpenAI-compatible endpoint.
We ran Ollama in production on bare-metal ThinkPads with no GPU. Then we replaced it with vLLM. This is the honest account of what broke, what improved, and what the right tool actually is for CPU-only inference in a real cluster.
The standard advice is Bedrock, Azure OpenAI, or Vertex AI. But what about organisations that cannot or will not send their data to American servers? Here is the full picture.
Ollama is a great developer tool. vLLM is what you run when real users are waiting. Here is exactly why the difference matters — and what changes when you move from one to the other.
A layer-by-layer comparison between minicloud — a 5-node k3s cluster assembled from CNCF components — and OpenShift Container Platform (OCP) self-hosted. Same problems, different approaches: immutable nodes, control plane management, security posture, upgrades, and what each model teaches you.
HA on a self-managed k3s cluster is not one setting — it is six independent layers, each with a concrete trade-off. This post walks through every layer on minicloud, what was validated by a live chaos game day, and how the same concerns map to EKS, GKE, and AKS.