Extending HAProxy Unified Gateway: backend config snippets and the failure modes I found
What HAProxy Unified Gateway needed to run the cloudauth filter: a config_snippet patch, a JSON bug only a real cluster showed, and four controller issues.
HAProxy leastconn vs power of two: the 2019 test rebuilt in 2026
HAProxy load balancing benchmark: the 2019 power of two test rebuilt on Kubernetes, plus gRPC on JVMs. leastconn holds up, peak EWMA won only in a narrow case.
Peak EWMA in HAProxy: from a Lua prototype to a core patch
Peak EWMA load balancing for HAProxy: a Lua prototype, a test bed on Kubernetes, a small core patch, and what the measurements taught about random draws.
Sovereign AI Agents: Context Limits Decide Where Your Data Goes
An AI agent reaches the limits of its model sooner or later. Whether its data then leaves your own infrastructure is a decision that has to be made on purpose.
Sovereign AI Agents: How My Own Fallback Can Send Requests to US Providers
Building an AI agent on a frontier model is easy, operating it sovereignly is its own discipline. My LLM series on OVH MKS shows where sovereignty quietly leaks.
Sovereign AI Agents: Routing, Egress, and Agent Permissions in Practice
Three control surfaces for a sovereign AI agent on OVH MKS: model routing and fallback in LiteLLM, egress under Istio Ambient, and tool permissions of the agent.