Adrita Das

Hi, I am Adrita! I am interested in working at the intersection of machine learning, AI systems, and biomedicine, focusing on scalable and theoretically grounded learning architectures with efficient implementations. My research centers on long-context and sequence modeling using state-space and implicit convolutional architectures, as well as unified models for molecular, geometric, and spatiotemporal data. As a secondary research interest, I am also drawn to biological foundation models and multi-modal generative AI—systems trained on genomic, structural, imaging, and single-cell data that learn broad biological representations. I’m excited about how such generative models can predict, simulate, and design biological states, revealing how molecular perturbations shape cellular and tissue-level phenotypes.

Recent work.
Recently, I worked on efficient diffusion models for 3D molecular generation, focusing on reducing the computational cost of 3D molecule synthesis without sacrificing structural fidelity. In particular, I studied Directly Denoising Diffusion Models (DDDMs) and provided a principled reinterpretation using the Reverse Transition Kernel (RTK) framework, showing that deterministic denoising corresponds to a structured transport map from noisy to clean molecular states. This perspective explains why few-step, deterministic diffusion can achieve fast and stable inference. Building on this, I developed SE(3)-equivariant, state-space–based diffusion architectures that scale to large molecules, capture long-range dependencies, and outperform stochastic diffusion baselines in both efficiency and generative quality on the GEOM dataset. This work was conducted in collaboration with Prof. Barnabás Póczos and Prof. José Lugo-Martínez. For a detailed overview of my research projects and publications, see the Publications page.

I also enjoy engaging with mathematics at a slower, more reflective pace—reading short linear algebra miniatures and solving non-standard statistics problems from recreational math books. These explorations often clarify fundamentals and reveal elegant structures that formal coursework sometimes obscures. I plan to gradually upload my personal notes and problem walkthroughs here. Stay tuned for updates!

Feel free to explore my projects and technical blogs. If you share similar interests or have exciting ideas to discuss, don’t hesitate to reach out via email. Let’s collaborate and create something amazing together!

What Languages Do I Use?

Python, R, C++, C, Java, MATLAB, CUDA

Technical Blogs

I enjoy consolidating my learning by writing blogs and short reports on mathematics and machine learning papers. Writing helps me clarify ideas, connect concepts, and reflect deeply on what I’ve read, making it an important part of how I continue to learn and grow as a researcher.

Understanding and Fixing Bottlenecks in State Space Models: What Recency and Over-Smoothing Tell Us

Sequence Models as Matrix Mixers: A Unifying View Beyond Attention

Click here to learn more about my:

Contact Me!

Thanks for visiting!!