A multi-agent system for academic paper discovery that learns research interests over time. Built using LangChain ReAct, LlamaIndex, and ChromaDB. Three specialized agents work together: one fetches papers precisely matching the topic (arXiv), one surfaces foundational context (Semantic Scholar), and one builds a persistent vector-map of interests using exponential temporal decay to connect current topics with past sessions.
Evaluated the robustness of Vision State Space Models against 21 image transformations (~15K images per transformation). Benchmarked VMamba-Base against Swin-Transformer and ConvNext, demonstrating consistent outperformance. Fine-tuned U-Mamba for road segmentation achieving >80% Dice score with custom gradient-freezing.
Analyzed how domain-specific knowledge is encoded in multimodal models like LLaVA and InstructBLIP. Implemented Domain Activation Probability Entropy (DAPE) and Logit Lens probing to track hidden representation evolution, finding specialized internal processing pathways for document understanding tasks.