Towards Blind and Low-Vision Accessibility of Lightweight VLMs and Custom LLM-Evals