My work sits at the intersection of deep learning research and practical ML systems. I’m especially interested in making training more reliable (optimization theory), models more useful (multimodal and graph learning), and evaluation more honest.

Research focus

Vision-language for medical imaging

At Symviq, I worked on vision-language modeling for chronic disease classification using retinal fundus images. I focused on CLIP-style contrastive objectives so that clinical, human-readable descriptions could act as supervision and improve generalization.

Typical themes:

• CLIP/SimCLR-style encoders for medical imaging
• Text-conditioning and multimodal embeddings for downstream classification and retrieval
• Foundation-model-style pretraining across multiple imaging tasks

Graph representation learning

My paper Graph State Networks (GSN) introduces persistent, nodewise selective state-space models for graphs — keeping per-node state across propagation steps as a scalable alternative to deep message passing. Published in TMLR, 2026.

See: Publications · Code.

Theory & optimization

My recent work studies asynchronous stochastic gradient descent under the Polyak–Łojasiewicz (PL) condition — a useful middle ground between convexity and general nonconvex objectives.

See: Publications.

Collaborate

If you’re working on multimodal learning, graph ML, or optimization theory for deep learning, I’m open to collaborations. The fastest way to reach me is by email.