GSoC 2026 interest in Project 10 Kubeflow Dynamic LLM Trainer Framework

[Pravallika Vetapalem]

Hi everyone,

I’m Pravallika. I’ve been following the Kubeflow project for a bit and recently graduated (2025) with a degree in Data Science. I’m really excited about the GSoC 2026 ideas, especially Project 10: Dynamic LLM Trainer.

Coming from a Data Science background, I've seen how fast the LLM space moves. I love the idea of making the Trainer SDK "future-proof" so it can easily support whatever new framework comes out next month (like Unsloth or Llama-Factory) without us having to rewrite the core client every time.

To get my hands dirty, I’ve already put together a small Proof of Concept (PR #310). I’ve been experimenting with a decorator-based registry pattern to handle backend discovery dynamically. I just pushed some updates based on the initial automated feedback to make it a bit cleaner.

You can check out what I’ve been tinkering with here: https://github.com/kubeflow/sdk/pull/310

I’m still refining my thoughts for the full proposal, but I’d love to hear if the maintainers think this registry-based direction is the right path for the project.

Looking forward to chatting more and hopefully contributing a lot this summer!

Best,                                                                                                                                             Pravallika Vetapalem