Portfolio

A self-programming AI agent framework for programmable workflows, visual desktop operation, and automated research.
openprogram.io

Persistent agent memory built from editable Markdown documents with references to immutable conversation records. Separate Writer and Reader roles organize and retrieve evidence.
GitHub

Training-free GUI grounding for off-the-shelf vision-language models. OCR and detected interface components provide coordinate references; iterative cropping refines the target before a click.
GitHub · Paper