Jupinder Parmar

Bio

I graduated with a B.S in Mathematics and M.S in Computer Science from Stanford University where I concentrated my studies within probability and machine learning. I was fortunate to be advised by Amir Dembo for my undergraduate degree and Christopher Ré for my graduate degree.

I now work as a Research Scientist at NVIDIA on the foundation model training team. I primarily focus on developing new large language models (LLMs) and augmenting their abilities via continued training.

// In the past, I've:

Worked with Tatsu Hashimoto on improving group robusntess of machine learning models. Current models achieve high average performance but can incur high error on certain groups of rare and atypical examples. We developed a new framework which allows for models to obtain better worst-group performance on tasks.

Worked as a teaching assistant for CS 234: Reinforcement Learning where I had the wonderful opportunity to help introduce the exciting field of reinforcement learning to students. Teaching has always been a passion of mine and it was gratifying to help aid in the educational journey of others.

Researched in Chris Re's Lab at Stanford where I worked with Khaled Saab. We worked on making use of passively observed human signals, such as one's gaze, to learn better representations of tasks and help train more robust, generalizable deep learning models for medical image classification.

Worked as a Machine Learning Engineer and the first Product Manager for OccamzRazor. OccamzRazor is an early stage company developing a biomedical knoweldge graph which aims to advance understanding of drug efficacy and provide novel treatments for diseases. My initial focus of work was on expanding out their natural langauge processing pipeline by researching and engineering new architectures for biomedical information extraction. Later, I focused on expanding the product platform from just repurposing known drugs to the broader set of small molecules that actively interact with any druggable protein.

Worked as a quantitative trading intern at IMC Trading on the Small Index Options Desk. I developed new trading strategies to improve performance on US Options Auctions by leveraging information regarding market sentiment.

Researched with David Chin at UMass Amherst where we investigated the intersection of quantum and classical machine learning. In particuar, we looked to see whether quantum models could help us better assess the quality of care provided by hospitals.

Worked with David Doty at UC Davis where we built a domain specific language that allowed researchers to implement complex theoretical models in the field of algorithmic self assembly.

Projects

Observational Supervision
Using passively observed human signals to improve medical imaging models.

Group Robustness
Improving model performance on atypical subgroups of data.

Biomedical Information Extraction
Uncovering biomedical data silos currently stuck within millions of pieces of literature.