I spend most of my time on the layer between a model and something someone will actually trust — tool-calling loops, retrieval, and the harness around them. By day that's agentic and RAG work inside Oracle OCI. Outside of that I work on kubeflow/docs-agent (GSoC 2026).
kmrrohit.space · github · linkedin · rr7433446@gmail.com · résumé
| where | what |
|---|---|
| kubeflow/docs-agent | GSoC 2026 — agentic RAG + MCP over docs, issues, code, manifests |
| Oracle OCI | catalog / order / planner agents · Alert Notification Microservice |
| MacBatch · app | batch inference on idle Apple Silicon, 252k embeds/hour |
| writing · projects · opensource | vLLM, SGLang, context, harness |
| merged | pr | note |
|---|---|---|
| Aug 15 | kubeflow/internal-acls#953 | org member |
| Aug 12 | kubeflow/docs-agent#232 | CI GHCR |
| Aug 08 | kubeflow/docs-agent#219 | session JWT |
| Aug 01 | kubeflow/docs-agent#218 | Helm gateway-guardrails |
| Jun 25 | kubeflow/docs-agent#210 | MCP + TEI + OKE CI/CD |
| Mar 09 | jaiakash/deploy-kubeflow#5 | Terraform OKE + Kubeflow |
- Teaching a docs agent to read the repo
- SGLang, or a runtime that remembers
- vLLM from the inside
- Context engineering, after the window stopped being the problem
- The harness is the product
- The two clocks
Local snapshot (public GraphQL — not github-readme-stats.vercel.app).
