Arbi Sahraoui

• Developed and deployed a real-time monitoring system for detecting Personal Protective Equipment (PPE) and workers and managing danger zones within video streams, for workplace safety. • Leveraged GPU acceleration and Triton server for enhanced performance. • Employed Apache Kafka for efficient processing of video streams. • Engineered an alerting service to quickly alert relevant stakeholders of safety violations. • Orchestrated Docker containerization to streamline deployment of microservices

• Created a tennis match analysis system for players, rackets, and ball detection and tracking. • Fine-tuned YOLO for players, rackets and ball detection, and exported the model to ONNX. • Implemented and trained a deep learning network integrating convolutional and deconvolutional neural networks, to precisely locate high-speed and small tennis balls surpassing YOLO by 60% in mAP achieving nearly 90% mAP. • Optimized model deployment with TensorRT, resulting in a 1.6x speedup in inference

Implemented a model for visual question answering that identifies and interprets temporal changes in time series images, enabling insightful responses to queries about evolving phenomena. • Integrated BERT for question embedding and ResNet101 for vision embedding and a fusion method to integrate textual and visual information. • Conducted experiments with a different setup, leveraging a visual transformer as the visual backbone achieving a 20% increase in accuracy compared to the ResNet101 configuration