Natural Language Processing (NLP) combined with Large Language Models (LLMs) represents the cutting edge of human-computer interaction through language. Our research focuses on developing advanced language understanding systems that can comprehend, generate, and reason with human language at scale. We work extensively with transformer architectures, fine-tuning pre-trained models for domain-specific applications, and implementing Retrieval-Augmented Generation (RAG) systems that combine parametric knowledge with external knowledge bases. Our RAG implementations enable more accurate, up-to-date responses by dynamically retrieving relevant information during inference. We explore multilingual capabilities, few-shot learning, prompt engineering, and efficient model compression techniques. Current projects include developing specialized LLMs for scientific literature analysis, creating robust dialogue systems for customer service automation, and building advanced text summarization systems that maintain factual accuracy while preserving semantic meaning.

Recommendation systems are crucial for personalizing user experiences across various digital platforms, including e-commerce and content streaming. Our research addresses the fundamental challenges of collaborative filtering, content-based filtering, and hybrid approaches that combine multiple recommendation strategies. We develop deep learning architectures, including neural collaborative filtering, autoencoders for implicit feedback, and graph neural networks that capture complex user-item relationships. Our work addresses critical issues, including the cold start problem for new users and items, scalability for millions of users, and fairness in algorithmic recommendations. Current projects include developing recommendation systems that incorporate temporal dynamics, building explainable recommendation models that provide transparent reasoning, and creating privacy-preserving recommendation systems using federated learning approaches.

Time series analysis is crucial for understanding temporal patterns and making predictions from sequential data across domains, including finance, healthcare, and IoT systems. Our research combines classical statistical methods with modern deep learning approaches to handle complex temporal dependencies. We develop sophisticated forecasting models using LSTM and GRU networks, Transformer architectures adapted for sequential data, and state-of-the-art models like temporal convolutional networks. Our work addresses challenges including non-stationary data, missing values, multivariate dependencies, and long-term sequence modeling. Current projects include developing real-time anomaly detection systems for industrial sensors, creating financial market prediction models with uncertainty quantification, and building energy consumption forecasting systems for smart grid applications.

Physical AI through Vision-Language-Action (VLA) models represents our cutting-edge research in embodied artificial intelligence that bridges perception, language understanding, and robotic control. Our research focuses on developing AI systems that can understand visual scenes, process natural language instructions, and execute physical actions in real-world environments. We are currently developing robotic AI foundation models in collaboration with Samsung Future Technology Foundation, creating systems that integrate computer vision, natural language processing, and robotic manipulation capabilities. Our VLA models combine multimodal learning approaches, enabling robots to comprehend complex instructions and perform sophisticated tasks in unstructured environments. We work with transformer-based architectures adapted for multimodal inputs, reinforcement learning for action policy learning, and imitation learning from human demonstrations. The models we develop aim to create general-purpose robotic systems that can adapt to various tasks through natural language instruction, representing a significant step toward intelligent physical AI systems.

Anomaly detection is crucial for identifying unusual patterns that deviate from normal behavior in various industries, including cybersecurity, manufacturing, healthcare, and finance. Our research develops both supervised and unsupervised approaches to detect anomalies in high-dimensional, complex datasets. We implement deep learning architectures including autoencoders for reconstruction-based detection, variational autoencoders for probabilistic anomaly scoring, and generative adversarial networks for learning normal data distributions. Our work addresses challenges such as imbalanced datasets where anomalies are rare, concept drift where normal behavior changes over time, and the interpretability of anomaly detection decisions. Current projects include real-time fraud detection systems for financial transactions, predictive maintenance systems for industrial equipment using sensor data analysis, and network failure detection systems.