Hi, I'm Aryan Bijva
I am an undergraduate researcher at GEC Patan, currently focused on the structural failures of foundation models when adapted for high-precision clinical tasks. My daily focus is a pursuit of mathematical alignment: specifically, how to prevent massive pre-trained transformers from collapsing into "templates" when they encounter the continuous variance of real-world medical anatomy.
Brief Overview
I am a 3rd-year Computer Science and Engineering student at Government Engineering College, Patan (Expected June 2027), maintaining a 8.74/10.0 CGPA. My academic trajectory is singular: I am preparing for a PhD at MIT EECS (Fall 2027). I have already secured national-level GPU compute grants from CDAC and am currently benchmarking novel geometric constraints for dense regression, moving beyond standard classification into the more rigorous domain of pixel-level medical metrics.
What Excites Me About Research
Geometric Deep Learning
Nature is not a grid; it is a manifold. I focus on Geometric GNNs to address the 76% variance collapse I identified in adapted foundation models like RETFound. By encoding physical coordinate priors directly into the latent space, I aim to align AI with the fluid, relational logic of biological structures and human perception.
My research is a pursuit of Truth over Templates. By "rethinking" adaptation, I move healthcare AI away from "black-box" guessing and toward manifold-aware architectures that preserve the individual biological variance necessary for personalized diagnostics. If we cannot recover the subtle geometric deviations in a patient, the model is not an assistant—it is a filter that obscures the truth.