What Kumail demonstrated

Built a perception validation dossier for a synthetic Aurora-7 detector with correct IoU-based detection rules, honest precision/recall scoring, and a safety-focused VRU false-negative analysis. Sliced recall by lighting, visibility, and rarity to surface the hardest pedestrian and cyclist cases, then mapped the real Uber ATG Tempe crash to the same night-time pedestrian failure mode using NTSB evidence.

• Leakage-Free Safety Metric Framing
• Condition-Sliced Failure Analysis
• Evidence-Based Incident Mapping