I am currently a Staff Research Engineer at Google DeepMind, focused on Reinforcement Learning Infrastructure for Gemini. I graduated with an MSc from McGill University, where I worked at both the Reasoning and Learning Lab and Mila.
I consider cycling to be the ultimate form of mobility in a city and advocate for better and safer routes for cyclists and pedestrians. Find me zipping around on my Brompton folding bicycle, looking for ways to live car-free in the Bay! Read about my journey on my blog, Bay by Brompton.
I love crafting narratives with different media to document my experiences. I perform improv whenever I can, a great way to stay present, practice active listening, and embrace the unexpected. I write a blog and graphic short stories as a way to share my point of view.
I have a long history of researching a diverse range of topics, from computational oncology to automated theorem proving. Most of my public research has focused on Reinforcement Learning, topics like policy gradient methods and affordances, and has been published at major machine learning conferences. Read about my published work on Google Scholar.
I am genuinely curious and always building new things both at work and personally. I hold a patent for hierarchical agentic methods for interacting with mobile devices and, am currently exploring new ways of interacting with intelligent language models that go beyond chat boxes.
I am currently part of a research engineering team that supports the development of Gemini, Google's family of multimodal AI models. I was involved end-to-end in the post-training of the Gemini 1.0 Pro and Ultra models focusing on SFT, RLHF, and eval/deployment. Since then I have been supporting research efforts aiming to further improve Gemini's capabilities by building infrastructure to unlock reinforcement learning ideas at scale.
I worked to develop the AndroidEnv library, an open-source environment for interacting with AndroidOS. This work was in service of pursuing hierarchical agentic systems that performed well on a challenging image-based environment with no OCR assistance. This work resulted in a patent
Exploring why entropy regularization helps policy gradient methods perform better by visualizing loss function landscapes. This work was done at Google Brain.
A technical discussion and tutorial. Completed during my internship at Datalogue.
Created a pseudointelligent chat bot on Facebook Messenger to aid students of McGill find information about courses, programs, buildings and printers faster.