This project looks at whether a group of LLM agents can reconstruct a person’s speaking style reliably.

System shape

The setup combines a discriminator and a mimic agent with automated prompt search. The agents iterate until the generated response becomes closer to the target style.

Evaluation

I evaluate performance using embedding similarity and qualitative checks on consistency, tone, and vocabulary.

Why I care

The project sits at the intersection of agent coordination, prompt engineering, and style modeling. It is also a useful testbed for thinking about how to make LLM systems feel more consistent across turns.