Researching dependable AI for Bengali and low-resource settings.

I build benchmarks and evaluation systems that make model behavior easier to inspect. Recent work asks whether retrieval systems find the evidence a Bengali user needs, whether multimodal models attend to diagrams and documents, and whether useful tutoring behavior can transfer to smaller offline models.

My undergraduate thesis focused on trustworthy AI and retrieval-augmented generation for low-resource agricultural intelligence. Across projects, I aim to connect each evaluation result to concrete system design choices and reusable public resources.