I am leading the AI team. My responsibilities include human resource management in the AI team, communication with project stakeholders, regular synchronization with the software development team, conducting internal team meetings.I implemented distributed GPU training of our Graph Neural Network + Diffusion model. I configured MLOps tools.I successfully deployed complex model (GNN+Transformer+Diffusion) into Triton Inference Server and provided SWE Team with HTTP/REST endpoints.Technologies: pytorch, vast.ai, pytorch, Google Storage, Google Cloud Build, Triton Inference Server

I was training models for Action Recognition (pick an item, put an item, etc) of customers in grocery shops, YOLOv5 for person and goods detection, Pose Estimation of customers, Unsupervised ReID of customers. I was solving the task of multi-object tracking. I was preparing data for annotation, annotation protocol. I was implementing custom tools for annotation. I was preparing models for deployment by converting them with TensorRT. I was configuring MLOps tools for the ML team.Tasks: Object Detection, Multi-Object Tracking, Action Recognition, ReID (Unsupervised), ClassificationFrameworks: yolov5, open-mmlab: mmdetection, mmclassification, mmaction2, mmpose.Technologies: pytorch, git, dvc, ci/cd, grafana, airflow, CVAT, aws s3.

I was conducting research on Image Editing using StyleGAN2, BigGAN, pix2pixHD models. The focus was on developing approaches of perturbating the model's parameters in order to get the image modified. The novel approach was presented at CVPR2021 Main Conference.Technologies: pytorch.

I was training an embedding model for fashion images using Metric Learning approach. Techniques used: Knowledge Distillation, Domain Adaptation, Angular Loss, ArcFace Loss. Improved version of embedding model is being used for solving Fashion Image Retrieval problem.I improved target matrics P1@1 from 41.2% to 48.4%. My model was being used in production.Technologies: tensorflow 1.x.