Rohan Potdar

I am currently on the Machine Learning and Software Engineering group at AMD, working on scaling the ROCm inference and training stack for models on AMD Instinct and consumer GPUs. I am interested in large language models and reinforcement learning. Previously, I graduated from Purdue Univeristy with a BS in Computer Engineering in December 2023. I also interned at Anyscale, working on LLM applications for Ray and offline RL for RLlib.

Email / Resume / Github / Linkedin

Research

I am passionate about artificially intelligent agents that can learn to reason and generalize over different tasks. Concretely, I would like to work on efficient algorithms for training and serving LLMs, model-based and offline RL, and applications such as recommender systems.

Outside of research, I enjoy playing the piano, squash, climbing, Brazilian jiu-jitsu, and exploring food across new cities!

Experience

	Software Development Engineer I'm on the ROCm AI team, where I'm working on an open-source benchmark to measure competitive performance of machine learning models on ROCm and Nvidia GPUs. I'm also consolidating internal CI/testing infrastructure to decrease build and test times across various hardware and model configurations.
	Research Assistant I researched robust control methods for autonomous surface vehicles under Dr. Shreyas Sundaram. Specifically, I worked on changepoint detection methods across Koopman-based neural networks and transformers on time series data to predict changes to the learnt dynamics model.
	Software Engineering Intern I worked on Ray, an open-source library for scalable distributed machine learning. In 2022, I implemented Doubly Robust Off-Policy Evaluation, a learning-based off-policy evaluation method for offline RL. In 2023, I worked on finetuning GPT-J with RLHF using TRLX and prompt engineering for code generation for the Ray open-source project.
	Teaching Assistant Fall 2021: Data Structures in C Fall 2023: Artificial Intelligence As a UTA, I taught recitation sections, held office hours, assisted with grading, and helped students implement projects and assignments.

Website template from Jon Barron
Last updated February 2024.