UniGPT Revisited: From a Simple Chatbot to an API-First AI Platform — Two Years of On-Premises LLM Operations

12 pages•Published: June 18, 2026

Jonathan Radas, Benjamin Risse and Raimund Vogl

Abstract

In 2024, we introduced uniGPT, an on-premises Kubernetes-based LLM platform at a major German university designed for GDPR compliance, digital sovereignty, and avoiding vendor lock-in. This paper evaluates nearly two years of operation (May 2024 - February 2026), tracing its evolution from a simple chatbot into a multi-modal, API-first AI infrastructure. Using the TOE framework, we analyze this progression as an iterative design cycle triggered by technological, organisational, or environmental factors. We detail 8 key iterations - including frontend and inference engine swaps, adding an OpenAI-compatible API layer, multi-modal services, and RAG pipelines. Notably, we find that >99% of usage now occurs via API rather than the chat frontend. Finally, we offer generalizable lessons for institutions building sustainable on-premises AI infrastructure in higher education.

Keyphrases: design science research, experience report, higher education, kubernetes, large language models, on premises ai, research

In: Laurence Desnos, Carmen Diaz, Janina Mincer-Daszkiewicz, Lazaros Merakos, Raimund Vogl, Stuart McLellan and Ulrike Lucke (editors). Proceedings of EUNIS 2026 Annual Congress, vol 109, pages 96-107.

Links:	https://easychair.org/publications/paper/wvMR
	https://doi.org/10.29007/4rq8

BibTeX entry

@inproceedings{EUNIS2026:UniGPT_Revisited_From_Simple,
  author    = {Jonathan Radas and Benjamin Risse and Raimund Vogl},
  title     = {UniGPT Revisited: From a Simple Chatbot to an API-First AI Platform — Two Years of On-Premises LLM Operations},
  booktitle = {Proceedings of EUNIS 2026 Annual Congress},
  editor    = {Laurence Desnos and Carmen Diaz and Janina Mincer-Daszkiewicz and Lazaros Merakos and Raimund Vogl and Stuart McLellan and Ulrike Lucke},
  series    = {EPiC Series in Computing},
  volume    = {109},
  publisher = {EasyChair},
  bibsource = {EasyChair, https://easychair.org},
  issn      = {2398-7340},
  url       = {/publications/paper/wvMR},
  doi       = {10.29007/4rq8},
  pages     = {96-107},
  year      = {2026}}

Download PDF Open PDF in browser