Piotr Trochim

Piotr Trochim

jolly fellow, science enthusiast, terrible skier

Career

  1. 2023 — present

    Deep Learning Entrepreneur

    Independent. Searching for a research bet to commit to. Working on life-long learning for LLM agents — how to make models that keep learning from their own experience after training, and how to keep their behaviour grounded as they do.

  2. 2017 - 2023

    Deep Learning Apprentice

    ML infrastructure and research at DeepMind and Meta. At DeepMind, tech lead across model-based RL, robot perception, causal reasoning, and the simulator infrastructure behind BCOOLER. At Meta, Uber-Tech Lead for the online experimentation infrastructure underpinning Ads Ranking.

  3. 2008 - 2016

    Embodied Intelligence for Games

    Animation, AI, and physics for games. Senior Gameplay/AI Programmer at CD Projekt RED on The Witcher 2; Senior Software Engineer at Havok building character and vehicle libraries used across the games industry; Crytek Frankfurt on the locomotion of the dinosaurs in Robinson the Journey (VR).

  4. 2005 - 2008

    Discovering Software Engineering

    First professional roles. Software Engineer at Motorola working on a Tetra-standard packet data router. Later at Sabre Holdings on next-generation data-mining architecture, mentoring teams on TDD and OOAD.

Projects I'm proud of

  • The Witcher 2 · CD Projekt RED

    Built the non-linear quest authoring system — a visual block language similar to Unreal Blueprints that orchestrated every gameplay system in RedEngine and was used end-to-end by writers and designers. Also designed the PathEngine ↔ animation integration, kinematic constraints for obstacle avoidance and docking, and engine systems including game-saves and world streaming.

  • BCOOLER — Industrial Cooling RL · DeepMind

    Hybrid simulator combining analytical solutions with multi-physics simulation, used as the training environment for DeepMind’s commercial-cooling RL system. The deployed policies cut energy use by ~9–13 % at the live experiment sites (deployment paper).

  • DeepMind Control Suite · DeepMind

    Open-source physics-based environments for continuous control, widely used as a benchmark in the RL research community. I built and open-sourced the visualizer for the suite, in desktop and web/Colab versions, and supported the user community.

Publications