Raj Vaingankar completed Machine Learning for Automotive Safety

Built a complete perception-validation dossier for a fictional Aurora-7 detector: defined detection at IoU >= 0.5, scored precision/recall and VRU false-negative rate from a synthetic detections table, and exposed a severe recall collapse in night and occluded VRU slices. The final recommendation was internally consistent and evidence-based, with a daylight-only release condition and a SOTIF-style analysis that mapped the Tempe crash to the same night-VRU functional insufficiency.

• Leakage-Free Detection Definition
• Honest Safety Metric Selection
• Per-Condition Failure Triage
• SOTIF Functional-Insufficiency Reasoning