Most of my open-source time is Kubeflow. As a GSoC 2026 contributor and Kubeflow org member, I built Flo, the agentic RAG docs assistant on docs-agent, and ran it in production with GitOps, evals, and user feedback. Before that, the Terraform/Helm path that puts a full Kubeflow platform on OCI.
Production pins a reviewed main commit and deploys the matching image tags, so a promotion is one small PR and a rollback is a revert. A script builds the promotion PR with changelog and eval results. Safe apps auto-sync through an app-of-apps, and a weekly CronJob trims Langfuse traces in ClickHouse and S3 (4 weeks live, 2 weeks eval).
Thumbs up/down on every answer, recorded as Langfuse scores against the trace, plus a nightly job that triages thumbs-down answers. Redesigned the docs chat widget in Material 3 with suggestion chips, dark mode, mobile layout, and a clear data-collection notice.
A staging copy of the agent that Argo CD keeps on main, with a PostSync hook that runs an LLM-as-judge eval set (Qwen) and posts results to Langfuse, so every merged change is scored before it reaches production.
Brought the whole deployment under Argo CD with multi-source Applications (public charts, private config), observe-only first so drift could be reviewed safely, plus the Qwen judge model release and session-issuer image builds in CI.
Markdown instructions and a label set that let the agent triage new GitHub issues consistently, grounded in the same retrieval as the docs assistant.
Improved retrieval quality across the docs, issues and code indexes and moved deployments to Helm upgrades, so releases are repeatable instead of hand-applied.
Fixed long agent answers being truncated at 30 seconds over A2A, and made replies grounded in retrieved sources with citations.
A small session issuer hands anonymous visitors short-lived JWTs that the Istio edge verifies, so the public chatbot is protected without asking anyone to log in.
Migrated public chatbot edge config out of raw Terraform kubectl heredocs into a Helm chart — Gateway, TLS, CORS lockdown, rate limits, and AuthorizationPolicies — so the edge is reviewable, versioned, and deployable with the rest of the stack.
Three Milvus-backed MCP tools, TEI embeddings, issues/code ingestion pipelines, Terraform for embeddings/KServe/Milvus/KFP pieces, and GitHub Actions that compile/test and optionally deploy to OKE (GHCR push + OCI kubectl).
Terraform modules for a full OKE stack (VCN, subnets, security lists, node pool) and a Kubeflow platform install (cert-manager, Istio, Dex, Knative, KServe, Pipelines, Central Dashboard, Profiles) with install scripts, guides, and troubleshooting for OCI.
The Kubeflow Docs Agent project runs a community call every other Saturday. It is a working session rather than a talk: what landed since last time, what is still open, where the priorities sit, and which pieces are small enough for someone new to pick up.
I am there as the Google Summer of Code contributor on the project, so most of what I bring is whatever the agent is currently getting wrong — a question its retrieval misses, an index that wants resplitting, a tool boundary that turned out to be drawn in the wrong place.
Anyone is welcome. Agenda items go into a shared doc ahead of time, and the Zoom link below is the standing one.