Návrh a implementace MLOps infrastruktury na Fakultě informatiky a statistiky

This thesis addresses the design and implementation of an MLOps infrastructure for the Faculty of Informatics and Statistics at the University of Economics, Prague. The motivation stems from a gap in the adoption of MLOps practices in the academic environment, as identified in the peer-reviewed literature, and from the specific operational needs of the faculty, which possesses a diverse portfolio of GPU cards without systematic resource isolation or observability. The thesis approaches the topic using the Design Science Research methodology and is structured into five sub-objectives: a synthesis of the peer-reviewed literature, an analysis of the current state based on semi-structured interviews with four faculty members, architecture design, implementation in a Microsoft Azure validation environment, and evaluation against the maturity framework by John, Olsson and Bosch (2021, 2025) and the catalog by Recupita et al. (2022). The implemented platform is built exclusively on open-source components. At its core are MLflow Tracking and Model Registry, Ollama with LiteLLM as the LLM inference layer, Caddy with TLS via Let’s Encrypt, JupyterHub with DockerSpawner, and the Prometheus monitoring stack with Grafana and dcgm-exporter. The infrastructure is declaratively managed via Ansible and Terraform. The assessment reveals a bimodal maturity profile. In the implemented components, the platform achieves Level 2 according to the primary framework, while in the components still in the design phase (data versioning, automated training pipeline), it remains at Level 0. Both issues identified in the analysis of the current state have been addressed—the lack of GPU isolation has been empirically resolved for the software path and designed for the hardware path with MIG, and limited operational visibility has been resolved with an observability stack. According to available sources, this work represents the first systematically documented MLOps implementation in the Czech higher education environment.