Arjun Dhamrait

Education

• University of California Santa Cruz 2024-present
  MS Scientific Computing and Applied Math
  adhamrai _at_ ucsc _dot_ edu
• University of California Berkeley 2017-2021
  BS Electrical Engineering and Computer Science
  adhamrait _at_ berkeley _dot_ edu

Research

• Lawrence Berkeley National Labs June 2025 - Current
  PI: Remi Lehe
  Investigated ML surrogate models that incorporate collective effects to replace computationally intensive steps in modeling particle accelerators. Added features to a framework for real-time guidance in laser-plasma experiments. Studying the scaling of backward-differentiable simulations of particle accelerators in the Cheetah open source project.
• UC Santa Cruz September 2024 - Current
  Advisor: Prof. Dongwook Lee
  Researched the use of Gaussian Processes for PDE solving and to update divergence-free fluxes of cells in Magnetohydrodynamics simulations.
• UC Berkeley, School of Information 2020 - June 2021
  PI: Xiao Qiang
  Assisted security research of popular proxy protocols (V2Ray/vmess, Trojan). Managed deployment of client testing servers within the GFW and servers outside on Digital Ocean. Helped prepare over 10GB of data for analysis via Neural Net. Thoroughly investigated protocol design and implementation for security flaws.

Industry Experience

• Meta July 2022 - January 2023
  Software Engineer
  Worked as a Software Engineer on a large, cross-functional team. Implemented features on a large internal react app to aid and assist sales, marketing, and business clients within Meta.
• Apple June 2021 - June 2022
  Software Engineer
  Served as the engineering point of contact of multiple large projects related to user-level OS features. Created and maintained a large cross-functional Swift utilities framework for use across Apple. Implemented features and fixed bugs in complex projects written in C/C++, Obj-C, and Swift.
• Apple June 2020 - August 2020
  Software Engineer Intern
  Migrated an existing cloud service onto Kubenetes. Created and maintained a CI/CD pipeline. Wrote a server implementing a new service API and deployed for internal use.

Projects

• Auto-Regressive Conditional Generative Model
  Implemented a generative model for multi-channel weather data using a conditional decoder. Modeled weather data using a variety of techniques, models, and architectures. Small model size enabled training for models accurate to 2 weeks in less than 5 minutes of training time and 30 seconds of prediction time.
  Report Code
• Fortran Linear Algebra Library
  Implemented many fast linear algebra routines for matrix decomposition, singular value decomposition, eigenvalue problems, and other matrix operations.
  Code
• De-sparsification via Variational Auto Encoder
  Implemented a desparsification model for multi-channel weather data using a VAE. Explored different models, hyperparameters, and architectures.
  Report Code
• Modeling High-Dimensional Lorenz-93
  Used neural nets to model a high dimensional Lorenz-93 system. Explored different methods of iterating steps of the model. Explored various implementations of a hybrid physics + ML model when modeling systems which have known physics.
  Report Code
• SinDY for Lorenz-63
  Used SinDY to extract the physics of the Lorenz-63 system with noisy exact derivative data and computed derivative data. Compared multiple methods of computing derivative from data, including finite difference methods, gaussian processes, and cubic splines.
  Report Code
• Optimal Gaussian Process Kernel Choice
  Used gradient descent to choose the optimal kernel function for a Gaussian Process. When approximating a sinusoidal function, The optimized kernel function weights obtained from maximizing the log-likelihood of the Gaussian Process correctly weighed the periodic kernel function considerably more than the other kernel functions. This choice of kernel function additionally allowed the model to predict unseen data very accurately
  Report Code
• Novel distributed large matrix inversion techniques
  Implemented multiple novel matrix inversion algorithms for large, multi-gigabyte matrices on distributed computers in the python package nums. Analyzed performance characteristics of different implementations onvarying matrix sizes.
  Report Code
• Calchart-4
  A typescript webapp for charting marching band shows
  Code