GSOC 2026

[Purushottam Jha]

Hi Edgar,

My name is Purushottam, and I'm a GSoC 2026 applicant interested in the project 'Expand kornia-vlm with ONNX and TensorRT backends.'

I want to be upfront: I made a mistake yesterday. I submitted PR #709 (a documentation PR) without first discussing it on Discord or linking a GitHub issue. Maintainer Christie Jacob correctly closed it. I've read the contribution guidelines and AI_POLICY.md carefully now — I understand the process and won't repeat that mistake.

About me:
- 1+ year experience with PyTorch, ONNX Runtime, and TensorRT
- Familiar with vision transformer architectures
- Rust beginner (but actively learning — built kornia-rs successfully and have been studying the codebase)
- Based in Bengaluru, India (IST timezone)
- No Jetson hardware access, but can use Docker/cloud for ARM testing

I've spent the last few days studying kornia-vlm deeply:
- All 3 models (PaliGemma, SmolVLM, SmolVLM2) use Candle exclusively
- The existing examples/onnx/ pattern using ort 2.0.0-rc.10 is clear to me
- I understand where ONNX Runtime needs to be integrated

Before writing any code, I'd love your guidance on:
1. Which model should be the first target for ONNX export — PaliGemma or SmolVLM?
2. Is a backend abstraction trait the right design direction?
3. What would be a meaningful first code contribution that aligns with your vision?

I'm committed to this project and will follow the community process correctly from here. Any guidance would mean a lot.
One more question: the problem/project actually mentioned in Opencv GSOC 2026, So I have to make a proposal and send it to the opencv ? or kornia, because in the problem Opencv attached the Kornia-rs repo link.


Thank you for your time.

Best regards,
Purushottam Jha