Research
Representations, adaptation,
and scientific discovery.
I study what learned representations preserve, how they change, and whether that information remains useful for downstream tasks.
01 · Representation & continual learning
When a model changes, what stays compatible?
My work with Atharva Airen examines representation drift in continual learning, feature alignment under uncertain correspondence, and compatibility in retrieval. These questions concern the relationship between a changing encoder and the systems that consume its features.
How can we distinguish a change in representation geometry from a loss of information needed by a downstream task?
02 · Foundation models, self-supervision & world models
Learning through prediction in latent space.
My ongoing foundation-model research focuses on learned representations: the information they retain, their behavior during adaptation, and their usefulness under distribution shift. I am interested in connecting self-supervised pretraining with continual learning, so that adapting a model to a new task does not discard information needed by other tasks.
I am interested in self-supervised learning and JEPA-based world models. My work with Aarav Gupta studies the distinction between pixel reconstruction and latent prediction, alongside action-conditioned sparsity for latent-space planning.
What does a training objective encourage a representation to retain, and which parts are useful for planning?
03 · Biology & molecular design
Biological and Molecular Representation Learning
I study how biological and molecular data can be represented to capture information relevant to disease, molecular properties, and chemical function. My interests include representations learned through self-supervision and pretrained foundation models, with applications in cancer research, drug discovery, and catalysis. A central question is how to adapt these representations when task-specific labels are scarce and assess their usefulness beyond the training data.
In molecular design, VECTOR+ combines a pretrained ChemBERTa encoder with property-guided contrastive learning, latent-space sampling, and molecular decoding. This provides a setting for examining how representation choice and adaptation affect generation from limited target-specific data.
In catalysis, my chemical ML work uses molecular descriptors, predictive models, and fragment recombination to prioritize phosphine-ligand candidates for cross-electrophile coupling.
My collaborative chemistry publications include methods for accessing pharmaceutically relevant pyridines from nitroarenes and for palladium-catalyzed distal C–H chalcogenation of aliphatic picolinamides.
Which information in a learned representation remains useful for new molecules or biological conditions when labelled data are limited?