Trajectory Graph Construction
Graphectory turns software-agent execution traces into directed graphs so repeated actions, reasoning loops, and tool usage patterns become inspectable instead of buried in raw logs.
Research Software and Visualization
Graphectory
Graphectory is a research codebase for turning software-agent trajectories into graphs and analyzing them through the taxonomy introduced in the associated paper. Instead of leaving those ideas as static figures, the project makes them explorable through interactive graph construction and trajectory-level metrics.
My role was on the software side: I led and implemented most of the browser-based visualization tooling that made the taxonomy usable in practice, and I contributed heavily to the release work that turned the lab code into a public repository.
Taxonomy-Aware Metrics
The system maps actions into phases like localization, patching, validation, and general work, making it possible to study how agents move through the taxonomy introduced by the paper.
Batch Analysis Pipeline
Precomputed graphs can be analyzed in bulk to produce trajectory-level metrics, while the browser tooling helps connect those aggregate metrics back to concrete runs.