Research

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

Current focus

Vision-language for medical imaging

At Symviq, I work on vision-language modeling for chronic disease classification using retinal fundus images. The direction I find most exciting is using CLIP-style contrastive objectives so that clinical, human-readable descriptions can act as supervision and improve generalization.

Typical themes:

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

Graph representation learning (GNNs)

I’m building GNN-Keras3, a Keras-first library that makes it easy to prototype GNN layers (GCN, GAT, differentiable pooling, etc.) with clean APIs and runnable examples.

See: Projects and GNN docs.

Theory + optimization

My recent research paper studies asynchronous stochastic gradient descent under the Polyak–Łojasiewicz (PL) condition, which is a useful middle ground between convexity and general nonconvex objectives.

See: Publications.

If you want to collaborate

If you’re working on multimodal learning, graph ML, or optimization theory for deep learning, I’m open to collaborations. The best way to reach me is via email (see the sidebar).