🔬 Research

My research focuses on Artificial Intelligence for Smart and Sustainable Built Environments, integrating sensing, prediction, reasoning, simulation, and intelligent control.

Research Interests

• Artificial Intelligence for Smart Buildings
• Occupancy Sensing, Measurement, and Prediction
• Large Language Models and Agentic AI
• Multimodal AI and Vision-Language Models
• Computer Vision for Indoor Environments
• Occupant-Centric HVAC Control
• Building Energy Management
• Model Predictive Control
• Reinforcement Learning
• Building Energy Simulation
• Time-Series Modeling
• Intelligent and Autonomous Building Systems

Occupancy Sensing and Prediction

A major part of my research investigates how building occupancy can be measured and predicted using real-world sensing data.

I have developed occupancy prediction approaches using multi-sensor data, Transformers, Mamba-based sequence models, and machine-learning methods. These studies consider both occupancy presence and occupant counts across multiple zones and prediction horizons.

The goal is to provide reliable occupancy information that can support adaptive HVAC operation and other occupant-centric building services.

Multimodal and Vision-Based Occupancy Measurement

My recent research extends occupancy measurement to surveillance video, object detection, multi-object tracking, large language models, and vision-language models.

These pipelines investigate how reasoning-enhanced AI can improve occupancy estimates when conventional computer-vision methods encounter occlusions, uncertain detections, and short-term fluctuations.

The estimated occupancy information is subsequently connected with occupant-centric HVAC control and building energy simulation.

Large Language Models for Smart Buildings

I investigate how large language models can be used for occupancy measurement, contextual reasoning, decision support, and intelligent building operation.

My work explores few-shot learning, in-context learning, structured prompting, reasoning mechanisms, and LLM-based interpretation of heterogeneous building information.

A central research question is how foundation models can complement conventional machine-learning and control approaches while maintaining reliability and constraint compliance.

Agentic AI for HVAC Control

My current research investigates agentic and evaluator-guided LLM architectures for closed-loop HVAC supervisory control.

These systems use building state information such as occupancy, indoor temperature, humidity, thermal comfort, and previous control actions to generate context-aware decisions.

My work also investigates deterministic validation and safety mechanisms so that LLM-generated recommendations remain compatible with physical building constraints.

LLM-Guided Optimization and Control

Another research direction combines LLM reasoning with established numerical-control methods.

This includes:

• LLM-guided model predictive control
• Multi-agent reasoning for occupancy, comfort, and energy objectives
• Reinforcement learning with LLM-guided action-space reduction
• Constraint-aware intelligent HVAC operation
• Robust decision making under uncertainty and agent disagreement

In these frameworks, LLMs provide contextual intelligence while numerical optimization or reinforcement-learning components retain responsibility for physical control decisions.

Building Energy Simulation

I use EnergyPlus and OpenStudio to evaluate intelligent control strategies under realistic building conditions.

Simulation studies connect occupancy information, indoor environmental conditions, HVAC operation, energy consumption, and thermal comfort to evaluate control performance.

My research aims to bridge the gap between AI prediction performance and practical building-level energy and comfort outcomes.