“The game of games had developed into a kind of universal language through which the players could express values and set these in relation to one another.”
— Hermann Hesse
Amirali Abdullah
AI Researcher · Mechanistic Interpretability · Representation Geometry
I am an AI researcher at Thoughtworks focused on mechanistic interpretability, activation steering, and representation geometry in foundation models. My work explores how internal model structure can be understood, measured, and leveraged for safer and more controllable AI systems.
A central theme in my research is the geometry of learned representations: how features are organized in activation space, how steering vectors interact with this geometry, and whether the resulting structure is consistent across models. I am particularly interested in non-linear approaches to representation steering and the alignment of geometric structure across model families.
I collaborate across academia and industry on interpretability tooling, psychometric evaluation of LLMs, and multimodal representation analysis. Previously, I worked in theoretical computer science on nearest neighbor search, dimensionality reduction, and information-theoretic geometry — problems whose core ideas around metric structure and spectral methods continue to inform my current work on feature spaces in neural networks.
Outside of research, I make a comic strip with my daughter explaining AI concepts to kids.
Affiliations: Thoughtworks
