The objective of the proposed work is to design scalable systems capable of summarizing biomedical literature, reasoning over curated biomedical knowledge, and generating interpretable hypotheses for complex clinical scenarios. Scientific AI assistants are becoming increasingly important for supporting biomedical research and clinical decision-making. However, the rapid growth of biomedical literature and patient data has made it difficult for clinicians and researchers to efficiently synthesize relevant evidence and derive actionable insights using traditional approaches. Existing clinical decision support and retrieval-based systems primarily rely on text matching or embedding similarity, which limits their ability to connect patient-specific information with deeper biomedical knowledge and mechanistic reasoning. As a result, these systems often provide limited support for hypothesis generation and treatment outcome reasoning. This proposal investigates the development of multimodal, retrieval-augmented biomedical AI assistants built on lightweight open-source large language models (LLMs). The proposed research aims to: (1) elucidate the current summarization capabilities of lightweight LLMs for biomedical literature, (2) develop a reliable and lightweight biomedical reasoning agent, and (3) develop a single- or multi-agent biomedical framework. Together, these components advance the development of interpretable and deployable biomedical AI assistants.