Research focus.
I used the HUD SDK to investigate how reliably an evaluation measures an agent’s ability to complete a task. My work connected evaluation environments, grader validation, and trajectory analysis: checking that task instructions, available tools, and scoring expressed the same objective.
The central research problem was measurement validity. A score becomes useful when the evidence explains what produced it, which alternative explanations have been tested, and how far the conclusion can reasonably extend.
Evaluation methodology.
Validate the experimental setup
Inspect environment behavior and tool interfaces, check consistency between task definitions and execution, and verify that the recorded evidence corresponds to the setup being evaluated.
Interrogate the reward signal
Reproduce scoring calculations independently, examine how the grader represents the task objective, and use controls to test for misleading signals and shortcuts.
Explain agent behavior
Analyze trajectories and tool failures, compare plausible explanations for success and failure, and distinguish agent limitations from problems in the evaluation itself.
Research contribution.
My contribution was connecting implementation-level debugging with experimental judgment. I investigated failure modes, checked reproducibility, and challenged whether a numerically correct result supported the behavioral claim attached to it.
This work brings together the parts of agent research I care about most: precise questions, inspectable experiments, independent checks, and conclusions calibrated to the evidence. Validating a score, validating the behavior it represents, and establishing generalization each require their own evidence.
Tooling.
HUD provides an SDK for defining agent environments and running evaluations. Its public repository documents the tooling used in this work.