My research designs interactive systems and models that make generative AI steerable. Today, people mostly steer AI with text prompts, describing what they want without seeing the range of what is possible. Instead, I build AI steering interfaces that turn a model's internal representations, such as its latent space and learned concepts, into controls people can see, explore, and directly manipulate. I study this through the lens of design, where goals are often ill-defined, take shape through exploration, and are hard to put into words.

Several questions I am exploring include:

• Latent Manipulation. What would direct manipulation look like for the latent spaces of AI models, and can the same operations generalize across different models and modalities?
• Steerable Models. How can we train models whose internal representations align with the abstractions people already use to think and create?
• Generative User Interfaces. When AI can generate interfaces in real time, how should they adapt to people's tasks, goals, and expertise?
• Steering Agents. When people delegate tasks to AI agents, how can they verify the agents' work, and when should agents surface decision points for human judgment?