Coding Google's PaliGemma VLM from Scratch
Reimplementation of a multimodal VLM stack from core components.
Implemented core modules for PaliGemma-style multimodal modeling, including the SigLIP vision encoder path, KV-cache flow, and Gemma language model integration with multimodal projection.
- Focused on end-to-end architectural understanding and reproducibility
- Used this project to benchmark reasoning and multimodal fusion behavior
- Placeholder: add full repo URL and benchmark table