Experiences

Projects

Dataset distillation compresses a large dataset into a small set of synthetic examples, but existing methods often specialize to a single task such as classification. We propose “drifting” synthetic examples through feature space toward the real data and away from one another. This produces datasets that better match the real distribution while remaining effective for classification, visualization, and other modalities.

Edamame gives each employee in a company an AI clone for organizational memory: we integrate Slack, Google Drive, Gmail, GitHub, Notion, and Jira with real OAuth and keep vectors in step as new messages, threads, and files show up. You query a clone in text or voice and get cited answers tied back to the underlying sources.