Annotation

Annotation: The Federal Treasury successfully operates 17 AI models, and the established distributed development process (prototyping at the RTU MIREA Research and Education Center, deployment at the Federal Treasury Data Center) has proven its effectiveness. Further improvement of this process requires addressing such limitations as the absence of a unified lifecycle management environment, loss of reproducibility, and high labor costs for manual model transfer. This defines the aim of the paper: to design and develop a unified on‑premise MLOps platform that ensures the full lifecycle of AI models with guaranteed identical performance in research and production environments. Main results: we propose a platform architecture comprising data preparation (MinIO, DVC), development and orchestration (MLflow, GitLab, Airflow), deployment (BentoML, Kubernetes, FastAPI), storage and caching (Redis, Kafka, PostgreSQL), and monitoring (Prometheus, ELK, Jaeger) modules. Scientific novelty lies in creating an integrated solution for a state financial authority that implements the "inner loop" principle and end‑to‑end automation of ML processes in isolated infrastructures. Practical significance is confirmed by achievable effects: reduction of model deployment time from 1–2 weeks to 2–3 hours, reduction of maintenance labor costs by 60%, and decrease of environment‑related incidents by 90%.