Maya Venkatraman

I'm Maya Venkatraman, a Research Software Engineer at the AlQuraishi Laboratory at Columbia University.

Currently, I work on developing a billion-parameter genome language model trained on prokaryotic DNA. My work spans the full ML research stack—from data curation and infrastructure to model architecture design, training, and biological benchmarking. I am particularly interested in the challenges of modeling biological sequences at single-nucleotide resolution and ultra-long context lengths, which push the boundaries of current transformer architectures.

In parallel to my research, I am pursuing a part-time Master's in Statistics at Columbia University, supported by the departmental MA2PhD Fellowship, which identifies students likely to pursue doctoral study in computational fields. I believe that rigorous statistical training is essential for advancing interpretable, efficient, and principled models in computational biology.

My interests within machine learning are both deep and broad, and I am considering multiple areas of focus for my PhD. I am particularly excited about the potential of diffusion models for protein design, as well as the use of reinforcement learning to guide exploration of conformational space. I am also drawn to novel AI techniques such as Hierarchical Dynamic Chunking from the Gu lab, valuing the idea of end-to-end, jointly optimized models that require less heuristic intervention. I believe that approaches like these will be especially relevant in biology, where data is inherently hierarchical, noisy, and context-dependent. I also see interpretability as a fascinating frontier in biological modeling, potentially enabling researchers to reverse engineer molecular contacts or mechanisms from patterns in model attention.

In the past, I worked at Google Research, applying computer vision to dermatologic image classification (Derm on Lens). I also contributed on a 20% basis to genomics projects, including DeepVariant and DeepNull. Before that, I worked at YouTube Trust and Safety, building infrastructure for detecting abusive user behavior.