• Developed NER model for different objections on speech outputs using PyTorch, resulting in an increase in accuracy by 25%• Created an automated dataset labeling tool using LLM prompt engineering and finetune• Developed a multi-GPU inference infrastructure for image labeling task, using Kubernetes, PostgreSQL and RabbitMQ, resulting in more than 1000 images/sec inference speed• Developed a multi-node training infrastructure for T2I task with grafana alert monitoring, using Kuberenetes, PyTorch distributed training and InfiniBand, resulting in more than 50 images/sec training speed

• Developed a transformer-based acoustic model using TensorFlow and deployed it in a production environment, leading to a 37% enhancement in speech recognition quality• Created a unified acoustic model capable of handling both streaming and non-streaming formats, resulting in a 1.5-fold acceleration in model training and the need to maintain just one model instead of two• Optimized model inference times through post-training quantization using TensorRT, enabling the deployment of a more resource-intensive model in the production environment. This enhancement led to a 53% improvement in speech recognition quality• Spearheaded the adoption of MLOps by implementing GitLab CI/CD and AWS workflows, resulting in streamlined development, testing, and deployment of ML solutions, accelerating the development process by 2-fold• Created and implemented a versatile BERT-based punctuation and capitalization model using PyTorch within a Kubernetes environment. This effort resulted in a 30% enhancement in language text processing efficiency• Utilized Named Entity Recognition techniques to extract valuable information from unstructured text data, resulting in an increase in accuracy by 30%