Built an end-to-end ETL and clickstream A/B testing framework to optimize checkout flows, reducing analysis time by 30%.
Processed ~200k–300k events for funnel/drop-off analysis and trained Random Forest & XGBoost models, achieving ROC-AUC 0.82 and 15% higher precision in conversion prediction.
Built a cloud-based AI pipeline using YOLOv6 on SageMaker to automatically detect traffic violations from city video feeds, improving detection accuracy by 30% and vehicle identification by 21%.
Automated video processing: Lambda triggers inference on new S3 uploads and aggregates results in real-time.
Added a GDPR-compliant anonymization plugin that masks faces and license plates using YOLOv6 and OpenCV, ensuring privacy while maintaining analytics accuracy.
Streamlined model updates and deployment with CI/CD, ensuring consistent and scalable operations.