Making multimodal AI more reliable.

I study robust visual representations and build useful systems that turn technical ideas into working tools.

Build the mental model.

Learning visual representations that hold up under noise.

Medical VQA systems often connect a frozen vision encoder to a language model. My work adds a denoising autoencoder before projection so the model learns to recover useful visual embeddings from corrupted inputs.

Conceptual pipeline from a noisy medical image representation to robust visual tokens
Noise-aware visual representation learning for medical visual question answering.
SLAKEBenchmark
PathVQABenchmark
LoRAEfficient adaptation
MLPVisual projection

Selected applications

Software for education, publishing, exam preparation, and public information.

Research should survive noise. Software should survive use.

I work across machine learning and software engineering, from controlled experiments to public-facing interfaces.