Vlad Petrutiu

- Responsible for decisions across the entire development lifecycle, from data management to algorithm selection and production deployment. - Developed a Zero-Shot Classification algorithm with 95% accuracy on over 2000 products in the Retail space which became the company’s core algorithm. - Explored efficient deployment techniques such as model pruning and model quantisation, utilising Docker as the primary containerisation platform. - Worked with CI/CD pipelines to streamline the integration and deployment of code.

- Explored various noise and blurring generation methods together with downsampling methods to create synthetic datasets for training Super Resolution algorithms. - Developed Transformer-based Super Resolution algorithms focusing on image reconstruction with emphasis on addressing noise, blurring, and 4x upsampling challenges. - Attained state-of-the-art results on various datasets, showcasing the effectiveness of the algorithm in comparison to existing methods. - Transitioned the project for potential deployment on phones, contributing to the practical application of the developed algorithm in real-world scenarios.

- Implemented a real-time solution utilising planograms and gap detection to identify missing products, streamlining inventory management processes in the supermarket using fixed cameras. - Successfully contributed to the development and deployment of a shelf-gap detection algorithm, leading to its successful rollout and integration into the company's services.

- Specialised in leveraging 3D objects, using Blender to model various food products for synthetic datasets used to boost the robustness of Computer Vision algorithms. - Collaborated with a dynamic team to bridge the gap between 3D synthetic data generation and real-world applications in canteen automation.

- Conducted research for my Bachelor's thesis, exploring direct applications of Machine Learning in Neuroscience. - Applied Wasserstein GANs to generate synthetic EEG signals, enhancing the accuracy of machine learning models in classifying real-world brain states. - Utilised GradCAM for thorough evaluation and interpretability, ensuring the reliability of synthetic EEGs and their impact on classification outcomes.