I was the Scoring and Grounding lead. I defined the rubric rules (pacing, pauses, pitch contour, energy), each with a feature, normalization, window, threshold and feedback template, and built the robust z-score (MAD-based) comparison that finds time-bounded deviation regions. I wrote the template explanations that name the measurement, the comparison, the time span and the rule ID, and the scoring that returns "insufficient evidence" rather than guessing. I also built the evaluation: IoU, boundary error, false positives and severity ordering on held-out speeches, plus a noise and compression stress test, and reported the failures honestly.
Log in or sign up for Devpost to join the conversation.