
The Machine Learning Toolkit for Kubernetes (AI)
Showing 5 of 30 projects. Click any project card for scope, mentors, and proposal studio.
Mentors: Student: Aniket Shaha
This project addresses limitations in Kubeflow Katib current Experiment CRD, which uses a generic and loosely typed interface for hyperparameter optimization. This leads to fragile parameter injection, inefficient resource usage, and unreliable state handling for stateful algorithms. To solve this, I propose designing and implementing a new OptimizationJob CRD v1alpha1, a strongly typed API focused on hyperparameter optimization for TrainJobs. It introduces native parameter injection, persistent algorithm state handling, and a push-based metrics pipeline integrated with the TrainJob Progress API. The project will deliver a production-ready controller with a structured reconciliation loop, support for stateless and stateful optimization algorithms, SDK integration, and comprehensive testing, resulting in a more reliable and efficient HPO workflow.
Mentors: Student: Krishna-kg732
The Kubeflow SDK gives AI practitioners a clean Python interface to submit, monitor, and manage distributed training jobs on Kubernetes via TrainerClient. However, LLM-based developer tools currently have no standardized way to access this runtime context — they cannot see a TrainJob's status, stream its logs, or reason about why a job failed. This project extends the existing MCP server MVP (tracked in kubeflow/community#936) for the Kubeflow Trainer SDK. The goal is to deliver a production-ready MCP server that exposes the full TrainJob lifecycle through well-typed, LLM-accessible tools — enabling AI assistants to actively help developers debug, monitor, and operate their Kubeflow training workloads in real time.
Mentors: Student: siddhant_jainn
Kubeflow's adoption at enterprise scale exposes critical bottlenecks in controller efficiency, security posture, and operational overhead. This project delivers six required improvements: (1) ConfigMap-driven database garbage collection for KFP pipeline runs, mirroring the existing artifact_retention_time pattern; (2) ScheduledWorkflow controller migration from legacy client-go informers to controller-runtime; (3) zero-trust NetworkPolicy enforcement for training job namespaces; (4) cross-namespace artifact security hardening with integration tests; (5) KFP dead code removal of deprecated manifests; and (6) Pod Security Standards compliance with CI regression gates. Stretch goals include Gateway API migration, Spark webhook readiness probes, and a Metacontroller migration design doc. This work builds on 12 pre-GSoC contributions (5 merged, 7 open) across 5 Kubeflow repositories.
Mentors: Student: danish9039
KFP and Katib users deploying via Helm currently have no upstream-supported path — charts exist but drift from Kustomize baselines silently, with no parity validation or explicit support boundaries. I've spent the last few months contributing to kubeflow/manifests (12 merged PRs) and reading through the maintainer feedback on PR #3237 closely enough to understand exactly what the constraints are. Before writing this proposal I built a PoC branch that gets all 9 scenarios (2 KFP + 7 Katib) to 0% parity diff against the upstream Kustomize baselines, traced and fixed each failure class, and ran runtime smoke tests on Kind. The goal is to upstream that work properly — charts, CI, documentation, and a support matrix that makes the supported/deferred boundary explicit. I also added two things that came out of building the PoC: a scheduled drift detection workflow and a schema validation gate using helm lint + kubeconform, both already implemented and validated.
Mentors: Student: Rohit Kumar (Kmrrohit)
Kubeflow's documentation, GitHub issues, and platform code are spread across dozens of repositories with no unified search. This project evolves kubeflow/docs-agent from a single-tool retrieval script into a multi-index Agentic RAG architecture with three MCP search tools (docs, issues, code), domain-specific ingestion pipelines (KFP), and a "Thin Context" MCP flow for IDE integration. All orchestration uses Kagent CRDs (Kubernetes-native). Deliverables: (1) Three MCP tools with dedicated Milvus collections and specialized chunking — issues at comment boundaries, code at YAML/AST boundaries; (2) Project's first test suite (71+ tests) with CI; (3) Idempotent pipeline upserts replacing the current destructive drop-and-recreate; (4) Developer IDE configs for Cursor/Claude Desktop; (5) Feedback logging for golden dataset accumulation; (6) RAGAS evaluation pipeline. Core work (PRs #140, #143) is already built and live-tested on OCI with 2,108 real chunks indexed.