• Trained SOTA multi-modal NFT spam detection system across 8 blockchain networks using fine-tuned LLM, GNN, XGBoost with Mixture-of-Experts ensemble, achieving 87% accuracy, correctly flagging 150,000+ malicious contracts

• Built trajectory-aware multi-agent pipeline with DSPy ReAct agents, integrating The Graph’s Token API and Subgraphs with judge-agent for evaluation; improved response accuracy by 70%, reduced query latency by 60% in production

• Designed RAG-based chatbot with BPS policy repository, automating document checks and reducing manual effort by 85%

• Aligned FAISS with vector and lexical search, metadata-driven clustering, and a multi-agent LLM pipeline, enhancing retrieval speed by 70% and document relevance by 90%

• Lectured and mentored ~90 graduate students in NLP, covering transformers, sequence models, and embeddings while guiding the implementation and evaluation of modern architectures

• Optimized and updated LERNet, AI4ALL, and Artemis websites, delivered lectures, and integrated a Makeblock robotics course to strengthen the curriculum with hands-on STEM and robotics learning

• Assisted in teaching CS506, guiding ~150 graduate students through core ML topics like regression, classification, clustering, decision trees, and model evaluation, while providing hands-on support in data preprocessing and algorithm implementation