WELCOME!

I am a principal researcher at Microsoft Research Cambridge, where I lead work on next-generation training algorithms and architectures for AI.

My broader research seeks to understand the computational principles of intelligence — and turn them into better ways to build AI systems.

(New!) Next-Gen Training Algorithms

Modern AI progress has been driven largely by scale. This project explores a complementary direction: better training algorithms that use compute more effectively.

One of our early results is the SinkGD Optimizer, developed in 2024 and recently adopted in DeepSeek V4.1 pretraining.

Building on this work, we develop ARO, a unifying view of LLM optimization through adaptive rotations — and a framework for designing more efficient training algorithms at scale. See Microsoft Research Forum Talk.

AI4Real-World Impacts

I also work broadly across probabilistic & causal AI for decision-making, with an emphasis on turning research into deployed systems.

One highlight is AI-driven personalized education: the technology I developed has reached more than half of UK schools. See media coverage: AI helps create personalized math lessons for students.

Bio

Before joining Microsoft, I did my Ph.D (2018- early 2023) in Machine Learning Group, CBL at the University of Cambridge, supervised by Prof. José Miguel Hernández-Lobato, and advised by Prof. Richard Turner. My PhD research focused on the field of probabilistic and causal machine learning. Check out my PhD thesis Advances in Bayesian Machine Learning: From Uncertainty to Decision Making. During my PhD, I also worked as an intern researcher at Microsoft Reserach Cambridge (MSRC), under the supervision of Dr. Cheng Zhang. Before joining the University of Cambridge, I obtained an MRes degree in Computational Statistics and Machine Learning from the Depertment of Computer Science, University College London, supervised by Prof. David Barber.