De Novo LLC (Ukraine)

•  Pioneer a national AI/ML cloud platform: Build Ukraine’s first cloud platform for AI/ML engineers, providing a ready-to-use workspace preloaded with all necessary software and tools to manage the machine learning lifecycle. The platform needed to eliminate the heavy lifting of infrastructure setup for data scientists, lowering the barrier to entry into AI/ML.
•  High-performance infrastructure: Implement a cloud environment accelerated by state-of-the-art NVIDIA tensor GPUs (H100 and L40S) to deliver unparalleled performance for training models and running AI workloads. This required integrating the latest GPU hardware with a robust virtualization and container orchestration stack to achieve world-class computation speeds for machine learning and inference tasks.
•  Data sovereignty and security: Ensure the solution is hosted in a reliable Ukrainian data center and complies with domestic security standards, so that government and enterprise clients can process sensitive data entirely within Ukraine. The platform needed to meet strict Ukrainian regulations (e.g. data protection by DSTZI certification) to enable AI projects for the public sector that cannot use foreign cloud services.
•  Cost efficiency: Offer AI/ML cloud resources at a significantly lower price point than global cloud providers (hyperscalers), with transparent and predictable pricing. The challenge was to optimize the infrastructure and pricing model to achieve up to a 2× cost reduction while maintaining high quality of service.
•  Scalability and multi-GPU support: Design the platform to support flexible scaling and multi-GPU workloads for large-scale modelsl. This included enabling distributed training across multiple GPUs and easy expansion of resources as user demands grow, all while providing integration with popular ML frameworks and orchestration tools (Kubernetes, Kubeflow, etc.).

•  Cloud platform architecture design: Designed the optimal architecture combining hardware (NVIDIA GPU infrastructure) and software components. This included planning the virtualization environment (VMware vSphere clusters) and container orchestration (Kubernetes via Tanzu) needed to support AI/ML workloads at scale, as well as incorporating NVIDIA’s AI Enterprise suite for GPU virtualization.
•  Infrastructure development and deployment: Built and configured the cloud environment for ML Cloud. This involved setting up the VMware virtualization layers, deploying Kubernetes clusters, and integrating the necessary network, storage, and security configurations in De Novo’s data center. We implemented GPU sharing technology (vGPU) to allow flexible partitioning of GPU resources (e.g. 1/8, 1/4 of a GPU) for different user workload. The team also established CI/CD pipelines for automating environment updates and ensured high availability and redundancy for critical components.
•  MLOps tools integration: Installed and integrated a suite of MLOps and data science tools into the platform. This included Kubeflow (for managing notebooks, pipelines, and model serving using KServe), MLFlow (for experiment tracking and model registry), MinIO (for scalable object storage of datasets and model artifacts), and annotation tools like CVAT and Doccano for computer vision and text data labeling. We configured these tools to work seamlessly together – for example, enabling Kubeflow notebooks to spawn with GPU access, linking CVAT to MinIO storage, and connecting MLFlow to Kubeflow pipelines – creating a unified ML workflow environment.
•  Testing and performance tuning: Conducted extensive testing of the ML Cloud platform. This included stress-testing with typical AI workloads (training neural network models, running inference on large datasets) to ensure the GPU scheduling and Kubernetes orchestration performed reliably under load. We fine-tuned resource allocation, optimized container configurations, and validated that multi-GPU training jobs and high memory workloads run efficiently. Security tests were also performed to verify data isolation between tenants and compliance with governmental standards.
•  Deployment and knowledge transfer: Collaborated with De Novo’s team to launch the platform into production. We prepared comprehensive documentation and provided training sessions for De Novo’s engineers on operating and supporting the ML Cloud environment. This knowledge transfer included how to onboard new customers, manage GPU quotas, monitor system health (using Grafana dashboards, Prometheus alerts), and troubleshoot common issues in the ML toolchain.