Built a risk-aware Visual Question Answering system by transforming BDD100K annotations into safety-focused Q&A pairs. Implemented distance estimation using geometric reasoning and fine-tuned BLIP-2 with QLoRA for efficient VQA training.

Built an LLM-driven pipeline that generates and iteratively improves Python Turtle drawings from real flower images using modular agents, visual evaluation loops, and diffusion-based refinement.

A Retrieval-Augmented Generation (RAG) system that is designed to search, summarize, and extract insights from academic research papers.