Kyle J Duffy

Research Engineer Pretraining Foundations @Cohere

London, GB

Doing whatever it takes to deliver powerful models.I pretrain Cohere\'s models to deliver new model capabilities to our customers. In addition to extensive data curation and model training, that consists of scaling our research infrastructure to handle new data formats, model architectures, and training algorithms. Most recently, that meant solving the modality alignment and engineering challenges that come with massive-scale multimodal VLM training.Before that I was a core member on the staff which delivered Command R7B Arabic, a best-in-class English-Arabic multilingual model we shipped in early 2025. I was the lead researcher responsible for model merging and candidate selection, as well as SFT and preference run planning.As we speak, I\'m working on shipping [REDACTED]

University of Oxford

Visiting Student, Mathematics

University of Notre Dame

Bachelor of Arts - BA, Mathematics and Computing

University of Oxford

Master of Science - MS, Advanced Computer Science