Case study: an AI assistance system at Hochschule Burgenland

Starting point and objectives

Higher education institutions face the challenge of using AI effectively without:

• relinquishing control of sensitive data,
• committing to proprietary SaaS platforms,
• or becoming dependent on individual providers over the long term.

The project therefore aimed to establish a modular, self-hosted AI platform that would:

• support different language models for different use cases,
• securely integrate internal knowledge,
• and allow flexible expansion of AI applications.

Technical implementation

The solution was implemented entirely on the institution's own infrastructure and tested under conditions close to production.

System architecture

• Operating system: Ubuntu Linux (LTS)
• Containerisation: Docker
• Reverse proxy: NGINX (TLS, routing, access control)
• Hardware: dedicated servers, including GPU servers
• Deployment model: self-hosted, with no shared cloud

All data and AI processing remain within the controlled system environment.

A multi-LLM approach: local and external language models

A central element of the project is multi-LLM operation.

Local language models (on-premises)

The following models were successfully connected on the institution's own GPU infrastructure:

• LLaMA 3
• Mistral 7B
• Gemini (local variant)
• gpt-oss:20b

These models process all data exclusively locally, without external disclosure.

External models (optional)

• ChatGPT through an API connection

External models are used selectively for less sensitive use cases or comparisons.

➡️ The model can be selected for each use case and changed at any time.

Knowledge-based chatbots with Mikis

Structured knowledge spaces (RAG)

Several workspaces were created with Mikis, each containing its own knowledge, such as:

• internal documents
• policies
• organisational information

Answers use retrieval-augmented generation (RAG):

• targeted use of stored content
• reduced hallucinations
• a traceable information basis

Embedding in websites

Individual workspaces were provided as website chatbots, for example:

• information chatbots
• service or support assistants

Internal and external use cases remain clearly separated.

A central AI frontend with the Mikis interface

Single sign-on and user management

• Microsoft Entra ID integration
• Authentication through OAuth2 / OpenID Connect
• Central single sign-on (SSO)

This enables controlled use through existing identity structures.

A consistent user interface

The Mikis interface serves as a central frontend for:

• different language models
• knowledge-based queries
• experimental and production AI applications

Advantages over proprietary SaaS platforms

Provider independence

• Open standards and open-source components
• No vendor lock-in
• Individual components can be replaced at any time

Full data sovereignty

• All data remains on the institution's own infrastructure
• No disclosure to third parties
• No shared tenant environments

Transparent costs

• Predictable hardware and operating costs
• Optional API usage only when needed
• No user- or feature-based licensing models

Extensive customisation

• Custom domain and corporate design
• Integration with existing SSO
• Tailored roles, chatbots and knowledge spaces

Long-term flexibility

• Modular expansion with new LLMs, tools and use cases
• Investment protection through an open architecture
• Independence from individual vendors or platforms

Project features at a glance

• Multi-LLM operation, local and cloud-based
• Institutional documents as a knowledge base (RAG)
• Embedded website chatbots
• Single sign-on through Microsoft Entra ID
• Self-hosted operation with Docker and NGINX
• Local image generation and image understanding using vision models