At the Cooperative AI Foundation, I developed a generative model of contextual integrity to study appropriate and inappropriate information sharing across social contexts. I designed and analysed experiments on how recipient roles, information needs, decision logic and dispositions affect sharing decisions.
At InstaDeep, I extended multi-task sequence models, prepared offline datasets and evaluated generalisation to unseen multi-agent tasks. I also contributed theoretical analysis and manuscript writing; this research was published on offline multi-agent reinforcement learning generalisation.
My research includes a first-author paper accepted at NeurIPS 2026 and co-first-author work on offline multi-agent reinforcement learning. I also extended the Sable architecture for multi-agent reinforcement learning across tasks and environments, and authored an online AI safety glossary.

