I'd rather ship something honest than something impressive.

I'm a data scientist and ML engineer, a master's student at Boston University, with a B.Tech in Computer Science from India.

My work has landed in three places that turn out to be the same problem wearing different clothes: generative models for medical imaging, where a class-conditional latent diffusion model synthesises the severe-grade X-rays a dataset barely contains; distributed systems, where I wrote KNN from scratch on PySpark RDD primitives rather than reach for MLlib; and applied AI, where a multi-agent retrieval system runs entirely on a local LLM. The common thread is scarcity — of data, of compute, of the labels you actually wanted.

Four of my papers are peer-reviewed, across IEEE Access, IJACSA and Springer Nature CCIS. That work taught me the habit I care most about: reporting the number that survived the held-out test set, not the one from the run that went well.

What I want next is a team where the model has to keep working after the notebook closes — where someone downstream depends on it, and the evaluation is honest because it has to be.