Utkarsh

Utkarsh

Ph.D. Candidate at MIT CSAIL

Massachusetts Institute of Technology

About

I am a Ph.D. student in Computational Science and Engineering at MIT, co-advised by Prof. Alan Edelman and Dr. Chris Rackauckas at the Julia Lab, MIT CSAIL. My research lies at the intersection of artificial intelligence, probabilistic modeling, and scientific computing. I build scalable, performance-engineered, and physics-consistent AI systems for scientific discovery, spanning LLM post-training, generative modeling, and high-performance computing.

Through my research internships, I have worked closely with Danielle Robinson, Bernie Wang, and Michael W. Mahoney at AWS AI Labs on probabilistic learning with hard constraints. At NVIDIA, I collaborated with Akshay Subramaniam and Noah Brenowitz on kilometer-scale climate emulation and, more recently, worked with DevTech on high-performance RL post-training infrastructure for mixture-of-experts language models.

Before MIT, I completed a double major in Electrical Engineering and Chemical Engineering, with a minor in Computer Science and Engineering, at IIT Kanpur, India.

Interests
  • AI for Science
  • Generative Models
  • ML Systems & HPC
  • LLM Post-Training & RL
Education

Publications

Selected work across AI for science and engineering, computational mathematics, and high-performance ML systems.

View
Loading publications…

Experience

Ph.D. Research Intern

Developer Technology (DevTech), NVIDIA Santa Clara, CA

Built high-performance RL post-training methods and GPU kernels achieving bitwise true-policy for MoE LLMs in Megatron-LM and NeMo-RL on NVIDIA Blackwell GPUs. Mentors: Daniel Galvez and Aurelien Chartier.
Now

Research Assistant

Julia Lab, MIT CSAIL Cambridge, MA

Research on LLM post-training with physics-based rewards, geometry-aware neural surrogates, physics-constrained generative models, and high-performance GPU differential-equation solvers. Co-advised by Alan Edelman and Chris Rackauckas.

Research Intern and Consultant

Julia Computing Inc. Remote

Developed pseudo-transient methods for differential equations with CUDA support and automatic differentiation. Mentor: Chris Rackauckas.