I develop machine learning methods for learning meaningful and controllable representations of the world, robust to noise, bias, and incomplete supervision. My work lies at the intersection of representation learning, multimodal modeling, and causality, guided by a central question:

"What makes a representation aligned with human conceptual understanding and amenable to control?"

I combine theoretical modeling with empirical analysis of vision–language systems to design interpretable, reliable representations that move beyond scale alone.

Current directions include:

• Causal and counterfactual modeling of multimodal signals
• Mitigating bias and misalignment in cross-modal training
• Language-guided identifiability and semantic control

My long-term goal is to enable trustworthy, human-aligned AI systems through principled representation learning.