Founding researcher at Geodesic Research (formerly ERA Cambridge). Developing techniques for building robust alignment initializations and removing dangerous capabilities from AI models through pretraining and midtraining interventions.