About

I’m Ilya Nyrkov. I build low-level compute clusters and deployment pipelines for hyper-scale machine learning systems. Currently, I am an HPC Researcher at the Technical University of Munich (TUM). Previously, I was a Lead Cloud Architect at VK Cloud.

Past Projects

• I engineered and maintain ozablas, an open-source, multiplatform C++ library that implements Ozaki Schemes to solve the precision-throughput divergence on modern AI accelerators. It achieves a 2x throughput speedup over native vendor FP64 libraries (rocBLAS). The AMD ROCm implementation was refined in direct collaboration with AMD engineering.

• I led the architecture and execution of a massive OpenStack SDN Migration. I designed the tooling and transition strategies to migrate VIP enterprise clients across entirely different software-defined network backends, successfully migrating over 200,000 live network ports while minimizing tenant downtime.

• I architected the largest multi-tenant Kubernetes cluster in the CIS cloud region (7 masters, 150,000+ objects). By bypassing standard KaaS bottlenecks, the cluster successfully scaled to handle CRM workloads of 20,000 active users growing by 1,000 users per day.

Contact

You can find my code on GitHub or connect with me on LinkedIn or contact me via email: ilya.nyrkov@gmail.com