Roman Vorushin

Research Engineer — More Intelligence per Compute Unit
Zurich, Switzerland·roman.vorushin@gmail.com·+41 77 426 32 71
LinkedIn·vorushin.github.io
Research Engineer with 13+ years at Google/DeepMind. Currently on Gemini, working across pre-training and post-training orgs. Deep expertise in JAX, Pallas, TPUs, and distributed training at scale. Getting more intelligence and research signals per compute unit.

Experience

Research Engineer at Google DeepMind — Gemini 2024 – Present
  • Post-training: Implemented and optimized post-training algorithms: SFT, PO, RL. Made ablations cheap and fast.
  • Pre-training bridge: Initiated and led a cross-org collaboration with the Large-Scale Pretraining org — running experimental ladders and measuring how architecture and data variants affect downstream post-SFT and post-RL quality.
  • Infrastructure: JAX, Pallas, TPUs, training codebase from pre-training to post-training.
Software Engineer / Research Engineer at Google 2012 – 2024
  • Built adversarial detection systems using vision-language and multimodal LLMs. Moved into the RL group at Bard/Gemini. (2021–2024)
  • Progressed through every generation of Google's training infrastructure: from training AlexNet variants on CPUs (DistBelief) to GPUs (TensorFlow) to TPUs (JAX) — writing training codebases from scratch at each transition. (2012–2021)
  • Designed planet-scale software systems at YouTube. Trained SOTA models for YouTube Content ID.
Software Engineer at Grammarly 2010 – 2012
  • Hired as a Python expert. Developed core backend systems. Discovered a passion for ML/AI that redirected my career trajectory.
Co-founder & CEO at Moodbox 2007 – 2009
  • Co-founded and launched a "visual Twitter" (covered by Mashable). Built the product, managed the team, and handled fundraising.
Software Engineer / Team Lead at Kazkommertsbank 2003 – 2007
  • Led the Internet-banking engineering team. Built and shipped production financial systems.

Education & Publications

Pavlodar State University — Electrical Engineering and Automation (1997–2002). Graduated with highest honors (5.0 GPA).
3 patents and 3 defensive publications at Google.
Completed Stanford online ML (ml-class.org) and AI (ai-class.org) courses (2011), MIT Linear Algebra (Gilbert Strang), Modern Robotics Course 1 (Northwestern/Coursera, 2025), and others.