My research whilst falling under the umbrella of deployable AI can be broken down to four pillars:

Pillar 1: Learning from Imperfect and Limited Data
I look at how AI systems can learn effectively when data is incomplete, imbalanced, or scarce. My work explores methods that make models more data-efficient and uncertainty-aware, such as selective memory replay, zero- and few-shot learning, and conformal prediction. I am particularly interested in how models can generalize to unseen conditions or adapt with minimal supervision, enabling reliable performance in realistic, imperfect data settings. 

Pillar 2: Efficient and Sustainable Intelligence
I study how to make deep learning models more memory-, computation-, and energy-efficient without sacrificing performance. This includes designing architectures that reduce redundancy, developing curriculum-based training that speeds up convergence, and creating benchmarks that quantify energy–accuracy trade-offs. My goal is to build AI systems that are efficient by design and can be deployed on everyday hardware while minimizing their environmental footprint. 

Pillar 3: Robust and Trustworthy AI
I focus on understanding and improving the robustness of AI systems against uncertainty, attacks, and distribution shifts. This includes work on federated learning to ensure reliability across diverse data sources, as well as studies of adversarial and prompt-based vulnerabilities in large language and vision models. I am also exploring how multi-agent systems can detect and resist steganographic or jailbreak-based attacks, with the broader aim of building AI that remains trustworthy in open and adversarial environments. 

Pillar 4: Foundation and Applied Medical AI
I apply my research in efficiency and robustness to healthcare, particularly medical imaging and physiological signals such as ECG and EEG. My work focuses on developing foundation models that integrate multi-modal data for clinical decision support, while ensuring that outputs are interpretable and reliable. I aim to create medical AI systems that can be safely deployed in both advanced hospitals and resource-limited settings, improving access to diagnostic intelligence worldwide.