If you’re paying attention to tech news, everyone’s building AI applications. RAG pipelines, agent frameworks, LLM wrappers. I went the other way.
I used AI to build infrastructure.
Specifically, I spent the last few months building a control plane for Envoy, the proxy that powers service meshes like Istio and sits behind most modern API gateways.
Here’s what makes this interesting: I’d never touched Envoy’s xDS protocol before starting—complete novice. xDS is Envoy’s configuration protocol—a set of gRPC APIs (LDS, RDS, CDS, EDS, SDS) that let you dynamically configure the proxy without restarts. It’s powerful. It’s also dense protobuf that demands deep protocol knowledge.
The thing that surprised me most was just how complex xDS actually is. Not conceptually—the ideas are straightforward. But the implementation details, the edge cases, the way resources reference each other. I’d underestimated it.
So I leaned heavily on Claude Code. And when I say heavily, I mean 70%+ of Flowplane was AI-assisted. Not just autocomplete suggestions—actual architectural decisions, debugging sessions, test generation.
What emerged is a control plane that lets you configure Envoy through REST APIs instead of writing raw protobuf. Create clusters, routes, listeners, filters—all through JSON that Flowplane translates into xDS. It handles JWT auth, OAuth2, rate limiting, header mutation, and 13 filter types total. Multi-tenant, team-scoped resources.
The most interesting part is the learning engine. Point it at traffic flowing through Envoy, and it infers API schemas from observed request/response patterns. No documentation? Watch the traffic, learn the shape.
The contrarian bet here isn’t Flowplane itself—it’s the approach. While everyone races to build the next AI application, there’s a whole layer of infrastructure that’s underserved. Envoy has millions of deployments. How many open source control planes exist for it? Not many.
Open source matters. Envoy proved that a proxy could be a commodity. Control planes exist too—Istio, Kuma, Contour, Envoy Gateway—all open source, all free. But they’re Kubernetes-first. Envoy Gateway recently added experimental standalone support, but it’s still built around CRDs and the Gateway API. If you want to run Envoy standalone—as an edge gateway, on VMs, without K8s—you’re either writing xDS config by hand or paying for enterprise solutions.
I’m not sure if Flowplane is the answer. But I’m increasingly convinced that AI doesn’t just help you build AI apps faster—it lets a single person tackle infrastructure problems that previously needed teams.
What infrastructure is underserved in your stack? If you’re running Envoy outside Kubernetes, or just curious about the project, reach out on github or find me on Linkedin
Top comments (0)