# Joshua Gevirtz > Machine Learning Engineer at Deepgram Location: Washington, District of Columbia, United States Profile: https://flows.cv/joshuagevirtz Deep learning system engineer. ## Work Experience ### Staff Engineer @ Deepgram Jan 2020 – Present I help develop the core deep learning production inference system that powers our high performance, high accuracy ASR models. This involves: * Low-level tuning of algorithms to maximize throughput and minimize memory footprint * Systems-level work to optimize data flows in an HPC context * Implementing cutting-edge deep learning models so that they can be put into production * Benchmarking deep learning systems as well as testing them for accuracy * Troubleshooting production issues to maintain high uptime ### Software Engineer (Systems) @ Determined AI Jan 2019 – Jan 2020 | Toronto, Canada Area Full stack development for machine learning platform. ### AI Engineer @ Deepgram Jan 2016 – Jan 2018 | San Francisco Bay Area I support deep learning research efforts at Deepgram through the development and maintenance of critical hardware and software infrastructure. Additionally, I help develop some of our SaaS products. Below are some of the projects I have worked on: - Developed optimized GPU versions of key proprietary algorithms using CUDA - Built and managed GPU cluster used for training Deep Learning ASR models - Developed custom Python modules (written in C) to perform fast GPU-to-GPU data transfers using OpenMPI during multi-GPU deep neural network training in Tensorflow - Wrote web application back-ends (SaaS product, system monitoring, etc.) using Python with Tornado and PostgreSQL and front-ends using Vue.js - Implemented custom Allreduce methods using OpenMPI primitives to optimize deep neural network model training - Developed Python data augmentation tools for audio data used in automatic speech recognition training tasks ### Freelance Software Engineer @ Self-Employed Jan 2015 – Jan 2016 | Berkeley, CA - Implemented novel audio search algorithm for client to maximize speech search performance. Delivered C++ code and Python wrapper. - Developed educational technology Android application. Python/Tornado and MySQL were used for the back-end. AngularJS used for the front-end of a companion web application. - Developed a web-based system for synchronization and administration of tasks across a network of embedded systems for an art installation ### Software Engineer @ Interana Jan 2015 – Jan 2015 ### Software Engineer @ IXL Learning Jan 2013 – Jan 2015 | San Mateo, CA - Expanded IXL's educational curriculum by implementing new practice modules using Java and JavaScript. - Developed new front-end framework using Facebook React. ### Graduate Research Assistant @ University of Michigan Jan 2012 – Jan 2012 | Ann Arbor, MI Working for the University of Michigan PandaX group required involvement in both hardware and software. The group's goal was to develop a prototype dark matter detector using liquid xenon. To this end, I developed a system to control a critical cryogenic gas-handling system. This included designing and building the valve-control hardware system, writing backend software in Python to talk to this system, and implementing a Python-based user interface to facilitate operation of the system. To support a grant-writing effort, substantial computational work had to be put into simulating the backgrounds of the experiment and the sensitivity of the detector to signal events. My contributions to this task are outlined below: -Expanded existing C++ GEANT4 simulation using NEST package -Wrote data analysis software to characterize detector response using ROOT C++ framework -Implemented neural network in Java to understand the technology's utility in event reconstruction -Studied optimal neural network training strategies using 48-core cluster ### Physics Laboratory Instructor @ University of Michigan Jan 2010 – Jan 2012 | Ann Arbor, Michigan Taught several sections of laboratory classes corresponding to calculus-based physics courses at the University of Michigan. Duties included discussing some of the background theory of the experiments, demonstrating equipment, assisting students with their work during the lab period, and grading lab reports and pre-lab quizzes. ### Graduate Research Assistant @ University of Michigan Jan 2011 – Jan 2012 | Ann Arbor As a graduate researcher in the ATLAS collaboration, I wrote analysis software to perform simulation studies useful in characterizing backgrounds at the Large Hadron Collider and helped develop software used for the Higgs Boson search in the ATLAS experiment. These responsibilities required heavy use of several C++ frameworks, as well as some use of Python for scripting and data visualization. Analyzing data for this experiment meant working with hundreds terabytes at once. Successful analyses therefore depended on highly distributed data analysis techniques - our analyses software was used on a cluster comprised of several thousand cores. My tasks are outlined below: -Wrote data analysis software in C++ using the ROOT framework for Higgs search -Wrote data visualization software using C++ and ROOT and performed Monte Carlo validation -Wrote Bash scripts to control the execution of analyses on CondorHT cluster -Wrote Python scripts to check data integrity and reacquire corrupt files from a central repository ## Education ### Master of Science (M.S.) in Physics University of Michigan Jan 2010 – Jan 2012 ### Bachelor of Science (B.S.) in Computer Science, Physics Ball State University Jan 2005 – Jan 2010 ## Contact & Social - LinkedIn: https://linkedin.com/in/joshgev --- Source: https://flows.cv/joshuagevirtz JSON Resume: https://flows.cv/joshuagevirtz/resume.json Last updated: 2026-03-22