Wiktor – AWS, Python, NumPy

The project aimed to develop an active learning system utilizing a vast unannotated data lake generated by a fleet of vehicles. Its primary objective was to autonomously identify significant data points for annotation and retraining, facilitating iterative enhancements of existing deep learning detectors and classifiers.

• Developed and implemented an active learning pipeline to automatically select and prioritize data samples from the unannotated observations lake for labeling;
• Integrated the system with existing deep learning models to facilitate continuous learning and enhancement of model accuracy and reliability;
• Coordinated with data engineers and annotation teams to streamline;
• Ensired the data flow and processing, significantly reducing the time from data capture to deployment;
• Utilized cloud technologies (AWS) and Docker to manage the scalability and operational aspects of the active learning system.

The project focused on implementing vehicle detection and multi-object tracking (MOT) models onto the NVIDIA Jetson NX platform, forming an efficient edge computing solution. It aimed to minimize latency and reduce dependency on cloud systems for critical perception tasks by enabling real-time processing directly on vehicles.

• Developed and optimized multi-task machine learning models for deployment on low-compute edge devices;
• Engineered a C++ inference application to run efficiently all models on the NVIDIA Jetson NX platform;
• Implemented robust multi-object tracking algorithms to enhance real-time response in varying traffic conditions;
• Spearheaded the deployment process, including setting up Docker containers for scalable and reproducible application management, monitoring, and adapting client applications to use real-time predictions.

The mobile perception platform that enhances safety, efficiency and sustainability. The project was related to the development of deep-learning models for Lane Layout Prediction in urban scenes

• Designed and prototyped multi-task deep learning models focusing on semantic segmentation and lane layout;
• Conducted extensive experiments and optimizations to refine model accuracy and performance under various urban conditions;
• Developed a flexible design to adapt the perception stack to new customer's requirements;
• Collaborated closely with data annotation teams to develop guidelines and ensure high-quality training datasets;
• Led the integration of models into the broader perception system, ensuring seamless functionality with existing automotive technologies;
• Helped build the internal DL team and secure the company's first customers.

An AI-powered fleet management platform. An open-set face recognition system was a core component responsible for automatically assigning routes to fleet drivers. The platform was designed to scale to thousands of fleets and drivers with minimal human annotation input.

• Designed, built, and trained a 2nd generation of metric learning DL models, significantly boosting face recognition performance;
• Led the deployment of the new version of the face identification system to edge devices and cloud processors;
• Improved metrics monitoring system;
• Developed and deployed internal tools for model-assisted data annotation, allowing for fast curation of the internal datasets.