Raj Babu Das

I designed and deployed an open-weight LLM chat service across two clouds. An Azure VM runs a kubeadm control plane and a Streamlit chat page. A JarvisLabs GPU VM joins that cluster as the worker and runs vLLM, so the model stays on the GPU while the UI stays on the smaller control plane.

The two machines do not share a network, and Azure's public IP is only a NAT address, so I linked them with WireGuard and ran the pod network over that tunnel. On the GPU node, I installed the NVIDIA Container Toolkit, scheduled vLLM with a Kubernetes RuntimeClass and a GPU label, and mounted the model weights on a local volume. Pausing the GPU VM stops compute billing and keeps the downloaded weights.

I wrote the setup, pause, and resume flow, and laptop access to the cluster through an SSH tunnel, with the Kubernetes API left off the public internet.

I am the creator and primary maintainer of MCP Gateway, an open-source control plane for MCP servers that centralizes server registration, tool discovery, authentication, authorization, and access management across teams and applications.

I architected and developed the gateway as a Go service supporting both HTTP and stdio transports, with upstream authentication via API keys, bearer tokens, JWT, and Basic Auth, including environment-variable-based secret interpolation.

It provides scoped endpoints such as /mcp/dev-tools, endpoint-level rate limiting, API key-to-endpoint mapping, usage counters, audit logging, tool-name namespacing, and persistent tool discovery using SQLite with embedded migrations. I also built the spec-first OpenAPI admin API and embedded Preact-based admin UI, and delivered multi-architecture Docker images and a Helm chart for Kubernetes deployments with persistent storage.

The gateway implements the MCP JSON-RPC lifecycle (initialize, tools/list, and tools/call) and persists discovered tool catalogs across restarts, providing a consistent and centrally governed interface for MCP clients.

I am a maintainer of LitmusChaos (litmuschaos.io), the open-source chaos engineering platform for Kubernetes. I maintain the control-plane GraphQL server, the subscriber and event-tracker agents that run in target clusters, and the Helm charts for the control plane, agent, and experiments.

I started by writing chaos experiments and the Kubernetes jobs that run them: disk loss on AWS and OpenEBS, CoreDNS pod delete, and OpenEBS pool network loss, network delay, and control-plane pod delete. I converted several pumba DaemonSets into node-selected Jobs, and added Ansible lint and Trivy checks to the experiment pipeline.

On the platform, I worked on the path from ChaosCenter to a target cluster. That includes event tracking for Litmus 3.x, moving cluster credentials into agent Secrets, node selectors and tolerations on the delegate manifest, readiness probes for the GraphQL and auth servers, security contexts with a read-only root filesystem, and an unprivileged nginx image for the dashboard. I also tightened environment-variable handling in the GraphQL server and fixed the GitOps path in the event tracker.

Across the LitmusChaos organization, I have 399 merged pull requests, including 250 on litmuschaos/litmus.

Selected for Google Summer of Code 2020 with the Cloud Native Computing Foundation (CNCF) to contribute to the Prometheus project. I developed a local Kubernetes cluster provider using KIND for Prometheus test-infra/Prombench, and enabled Prombench deployments on both Amazon EKS and local Kubernetes clusters. I improved Grafana dashboards used for Prombench benchmarking and contributed enhancements to the Prometheus test infrastructure.

I worked extensively with Go, Kubernetes, Prometheus, KIND, AWS/EKS, Grafana, and open-source CI/testing workflows.

Selected as an LFX Mentee through The Linux Foundation’s LFX Mentorship Program, contributing to the Keptn project from March to May 2021. I worked within the Keptn open-source community, gaining hands-on experience contributing to a CNCF ecosystem project and collaborating through the project’s open-source development workflow.