Team lead for a team of 4 data scientists and machine learning engineers, responsible for the development of a multimodal (text, audio, classic) AI systems for real-time text and audio processing. Responsibilities: – Leading the team in the research, development and deployment of a multimodal AI system. – Mentoring and guiding team members in their professional growth. – Collaborating with cross-functional teams to ensure successful project delivery. Achievements: – Hired and mentored a team of 4 DS and MLE. – Built from scratch all the processes and infrastructure for the team, including: –– Model development and deployment processes –– CI/CD pipelines for model deployment –– Version control and artifact management –– Monitoring, logging and alerting – Under my leadership, the team has successfully developed and deployed: –– A multimodal AI system combining ASR, TTS, and LLM components for realtime phone conversations that achieve human-level performance. –– A comprehensive AI solution for automated phone conversation quality assessment and analysis

Development of a Text-to-Speech, Speech-to-Text, and ASR services that outperform leading competitors in the Russian language. Responsibilities: – Model deployment using Triton Server, Docker, Kubernetes, Python, and Golang. – Model optimization for performance and resource efficiency: – Model architecture changes – TensorRT, tritonlang, torchcompile and OpenVINO – Model warmup, quantization, sparsity and pruning – Research, develop, train and fine-tune new model architectures. Achievements: – 0.12 p95 latency and 90 RPS for diffusion model on a single 2g.20Gb A100 instance. – Established and automated a version-controlled model deployment process using CI/CD, WandB / ClearML, and Artifactory. – The development process was established following best practices (e.g., CI/CD, code review, documentation, unit testing, changelog, semantic versioning, etc.). – Created an automated quality and performance testing in a production-like Kubernetes environment. – Implemented comprehensive observability with monitoring, logging, and alerting. Technologies: Python, Golang, PyTorch, tritonlang, CUDA C++, torchcompile, TensorRT, WandB / ClearML, Triton Server, Observability, Docker + Kubernetes, Git + CI/CD

Designed and implemented machine learning service for the automatic opening of car barriers using computer vision. Achieved 99% accuracy in recognizing Russian license plates and 90% accuracy for all other license plates, maintaining a latency of 0.5 on a 4-core CPU. Technologies: Python, PyTorch, OpenCV, Docker + Kubernetes, OpenVINO + Triton Server, Observability (Grafana, Kibana, Prometheus), Git + CI/CD

Designed and implemented service for building facade segmentation, managing everything from data collection and preprocessing to model training, evaluation, and deployment. Technologies: Python, TensorFlow + Keras, OpenCV, Docker, Git

Designed and implemented facial recognition service for employees, covering everything from data collection/generation and preprocessing to model training, evaluation, and deployment. Technologies: Python, TensorFlow + Keras, RealSense DepthCamera, OpenCV, Docker, Git