An interpretable approach to language-model reasoning that searches over explicit, high-level reasoning actions. I co-authored the work, accepted to COLM 2026, to make generation more controllable and diverse.

My ongoing research with Zach Bamberger and Lakshya A Agrawal on improving prompts and agent skills from execution traces. The approach combines GEPA with a tool-calling editor to make and evaluate targeted changes.

An open-source, self-hostable platform I built to develop, optimize, and serve LLM programs. It turns datasets and custom metrics into DSPy programs, improves them with GEPA, and evaluates results on held-out data.