← Back to projects

HPC MAISTER — Local AI Platform for RCUM

I design and test an internal AI platform on university HPC infrastructure. My ultimate goal is a locally hosted AI assistant for internal use at RCUM, helping staff find information in authorized documentation and support their daily work.

Executive summary

I combine target architecture, an isolated identity-enabled pilot and reproducible local LLM evaluation. I treat permission-aware retrieval, cited answers and operational acceptance as explicit requirements rather than completed production capabilities.

Linux OpenWebUI Keycloak PostgreSQL OpenID Connect LLM Evaluation RAG Architecture

Technical architecture

Problem

  • I need to connect internal knowledge access with university identity and authorization boundaries.
  • I need to evaluate local model quality and resource use before selecting an operational configuration.

What I built

  • I designed a target architecture for permission-aware retrieval, source authorization, verifiable citations, auditability and unsupported-answer handling.
  • I implemented and tested an isolated OpenWebUI, Keycloak and PostgreSQL pilot with HTTPS and role-based OIDC access.
  • I worked on Microsoft Entra integration testing and paired Slovenian/English evaluation of local LLM inference.
  • I prepared acceptance criteria, test scenarios and operational procedures for further integration.

My role

  • I work across architecture, Linux-based infrastructure, identity integration, evaluation design and technical documentation.
  • I record reproducible configurations, latency and throughput observations, quality results and known limitations.

Constraints

  • I keep internal documents, infrastructure details, identities and access configuration private.
  • I separate a functioning component pilot from complete retrieval, authorization and production acceptance.

Outcomes

  • I established an implemented component pilot and a documented basis for further platform decisions.
  • I made model quality, identity behavior and integration acceptance explicit parts of the work.

Architecture decisions

  • I keep permission-aware retrieval and source citations in the target design.
  • I evaluate model performance alongside bilingual answer quality.
  • I require explicit integration and operational acceptance before treating a pilot as a deployed service.

Public evidence

I describe my role, the implemented pilot components and the evaluation approach in this case study and my RCUM experience record. I keep execution records and internal artifacts private.

Proof limits

I present architecture and implemented component pilots. I do not claim a production AI service, completed end-to-end RAG acceptance, access to all internal documentation or measured organization-wide productivity gains.