Research

Specify, check, and revise

AI agents make many choices on people's behalf that people do not see. I make a person's intent explicit and shared with the agent, so people can:

  1. Specify what matters
  2. Check the agent's choices against it
  3. Revise it as things change

Each system here has a demo you can try in the browser.

PromptPath

Specify and revise

Keeps a persistent profile of the user and uses it to repair short prompts, showing which words came from the prompt, the profile, or the AI.

Try the demo

VeriPlan

Check and revise

Checks an LLM's plan against rules the user confirms, using model checking, and sends any violations back to the user and the LLM.

Try the demoPDF

U-Define

Check and revise

People state their own constraints in natural language and mark each one as hard or soft. Hard rules are checked with model checking, and soft preferences are rated by an LLM judge.

Try the demoPDF

MAP

Specify

Collaborative LLM-powered agents that personalize for several people at once and resolve conflicts among their preferences and schedules.

Try the demoPDF

Foundations for people and AI to communicate

I have long been interested in how people and AI systems communicate: designing the foundation for it, and understanding how AI plays out in real settings.

Toward embodied agents

My goal is to extend these mechanisms to embodied AI, as agents move into homes and shared spaces.