Welcome!

My research focuses on the mathematical and algorithmic foundations of trustworthy AI, with emphasis on privacy-preserving learning and inference, evaluation of model and agent behavior for alignment and safety, and memory- and compute-efficient inference for large language models and AI agents. From October 2024 to October 2026, I was a Postdoctoral Fellow in the Information Theory Lab at Harvard University, where I worked with Professor Flavio P. Calmon.  From December 2022 to August 2024, I was a Postdoctoral Researcher in the Biometric Security & Privacy group at the Idiap Research Institute, where I worked under the supervision of Professor Sébastien Marcel. I received my Ph.D. in Computer Science from the University of Geneva in 2022, under the supervision of Professor Slava Voloshynovskiy in the Stochastic Information Processing (SIP) group. During my Ph.D., I spent six months as a Visiting Research Fellow in the Information Theory Lab at Harvard University, where I collaborated with Professor Flavio P.  Calmon, and six months as a Visiting Research Scholar at Imperial College London, where I worked with Professor Deniz Gündüz. Prior to my Ph.D., I earned an M.Sc. in Mathematics (Numerical Analysis) from Iran University of Science and Technology in 2017, and an M.Sc. in Electrical Engineering (Communications Systems) from Ferdowsi University of Mashhad in 2014, under the supervision of Professor Touraj Nikazad and Professor Ghosheh Abed Hodtani, respectively.

Research

My research lies at the intersection of machine learning, information theory, and signal processing. I develop theoretical and algorithmic methods for measuring and limiting information disclosure, preserving task-relevant information, and characterizing privacy-utility trade-offs in learning systems. My research encompasses information-theoretic and differential privacy, representation learning, algorithmic fairness, and federated learning. My current work extends these foundations to generative AI and LLM-based agents, with particular attention to privacy leakage across prompts, retrieval mechanisms, intermediate representations, persistent memory, tool calls, interactions with external services, and generated outputs. I also study behavioral evaluation for alignment and safety, as well as the memory and computational costs of large-language-model inference.

My research is guided by the following questions:

• How can privacy-preserving mechanisms be designed and analyzed under explicit threat models and formal privacy guarantees for systems that process sensitive data, including biometric and digital-health systems?

• How can information leakage be measured and mitigated throughout language models and tool-using AI agents, including their prompts, retrieval mechanisms, semantic representations, persistent memory stores, tool calls, interactions with external services, and generated outputs?

• How can learned representations preserve task-relevant information while limiting leakage about sensitive attributes and satisfying explicit representation-complexity or computational constraints?

• How can disparities in model predictions and downstream decisions be identified and quantified, and how can they be mitigated while characterizing the resulting fairness-utility trade-offs?

• How do formal privacy constraints and statistical heterogeneity affect model utility, convergence, communication cost, and computational cost in federated learning systems?

• How can evaluation protocols quantify privacy leakage, calibration, policy compliance, and robustness to distribution shift, adversarial prompting, retrieval failures, and multi-turn interactions in language-model systems?

• What are the fundamental trade-offs among memory footprint, latency, approximation error, privacy leakage, and downstream task performance when inference-time representations in large language models are compressed, transmitted, stored, or reused?

These questions define a research program on trustworthy and resource-efficient AI, with applications in biometrics, healthcare, medical imaging, retrieval, and generative AI. My broader goal is to develop theoretically grounded and empirically testable methods for AI systems subject to explicit privacy, fairness, safety, reliability, and resource constraints. I seek to connect formal analysis with reproducible implementations and empirical evaluation in settings where failures can expose sensitive information or degrade system reliability.