My research interests lie in building emergent models of intelligence that are helpful, robust and aligned with human values. To pursue this, my current research direction is in the science of scaling large language models that are capable of complex instruction following and agentic behaviour, and which are human-aligned. My prior research also spanned the intersection of language + X ∈ {learning, representation, bias, reasoning} and technology + Y ∈ {ethics, culture, government, security}.

At Cohere, I was the technical lead for post-training recipes. I led a research team to determine the best post-training recipes and apply them to develop Cohere's flagship large language models that power our agentic AI platform for enterprises. Our Command {A+, A, R7B, R+, R} and Aya models are available as open-weights for the research community.

Apart from research, I also have experience in products and policy. At Waymo, I developed models to determine the true collision rate of self-driving cars in simulation given real and simulated driving episodes, for the purpose of accurate safety benchmarking. At Facebook, I used transformer-based models for the detection of harmful comments on Instagram. At Singapore's Smart Nation and Digital Government Office, I authored strategy papers and analyses on the societal harms of AI and the responsible deployment of human-centered AI systems.

At Yale, I was advised by Dragomir Radev (LILY Lab), Robert Frank (CLAY Lab) and Elisa Celis (Controlling Bias in AI Group). I also worked with John Lafferty to create a curriculum for, and teach, text data science.