Open Source
GraphBit is a graph-based agentic AI framework for non-linear multi-agent orchestration. Instead of chaining agents in a fixed sequence, it models agents, tools, and memory as nodes in an executable graph, enabling branching, parallel, and cyclic workflows. I lead the project and designed the graph-based orchestration model, the execution engine that schedules nodes across branching and cyclic paths, and the memory and tool interfaces that agents share. I also drive the open-source release and roadmap, building the framework for autonomous multi-agent systems that need dependable, inspectable execution rather than opaque agent chains.
AI for Cyber-Physical Systems
Ongoing Research · MTSU
This line of work applies graph neural networks, attention, and LLM agents to physical systems that must be monitored and controlled from sensor data: power grids, photovoltaic inverters, and autonomous drones. It includes a dual graph attention network for robust open-switch fault diagnosis in PV inverters, an unsupervised fault detection and classification scheme for power converters, and self-attention CNNs with time-series imaging for transmission line fault detection and localization. At MTSU I am modelling grid topology as graphs so GraphSAGE and graph attention networks can reason over interconnected buses, lines, and sensors, coupled with LLM reasoning for explainable decision support. In parallel, I am integrating LLM agents with onboard perception and control loops for high-level drone mission planning, using ROS, Gazebo, and Open-RMF for simulation and robotics tooling.
Hyperspectral Imaging
Research
A sustained line of work on deep architectures for hyperspectral remote sensing imagery, where each pixel carries hundreds of spectral bands and labelled samples are scarce. It began with a multidimensional pixel-wise CNN and a regularized SVD-based multidimensional CNN for compact feature extraction. It then moved to a spectral-spatial residual attention network and a dynamic kernel network with adaptive receptive fields. Most recently, it produced a unified framework of energy-based generative models with morphological attention networks for hyperspectral image classification.
Non-Convex Optimization
Ongoing Research
Training deep networks, graph neural networks, and energy-based generative models means solving high-dimensional non-convex problems where the loss landscape is riddled with saddle points and poor local minima. This work studies how regularization, low-rank structure, and probabilistic modelling shape that landscape and make optimization tractable and stable in practice. It draws on regularized singular value decomposition and regularized encoder-decoder generative probabilistic networks to constrain the solution space, and on energy-based models whose training requires optimizing intractable partition functions. The goal is optimization methods that are reliable under the noisy, imbalanced, and data-scarce conditions typical of sensor-driven and remote sensing applications.