Education / Sovereign AI

Sovereign Student Assistant: On-Premise RAG for Higher Education

Client A US education technology company

For a US education technology company, Jungroo AI built a self-hosted assistant that answers students' questions about admissions, financial aid and courses from a college's own web pages. It runs entirely on the client's GPU server, so student information never leaves its environment. A structured evaluation of five retrieval models and three language models chose the final stack, which scored 98.5% retrieval accuracy and 95.3% answer accuracy on our test sets.

The challenge

Students ask the same questions again and again, and the answers are spread across thousands of web pages. The client needed answers that were accurate, quick and grounded in the institution's own pages, with personal information kept inside its own environment instead of being sent to a public AI service. Keyword search over summarised pages missed context and handled follow-up questions poorly.

The solution

Jungroo AI designed, evaluated and deployed a retrieval-augmented assistant on the client's own hardware:

• Evaluation Framework: A repeatable test harness that compared five retrieval models and three language models on the institution's own content before the stack was chosen.
• Semantic Retrieval: Vector search over the full text of every web page, in place of keyword search over summaries.
• Follow-Up Understanding: Each follow-up question is rewritten into a complete query before retrieval, so conversations stay on track.
• Self-Hosted Models: Granite Embedding R2 for retrieval and Gemma 4 26B for answers, both running on the client's own GPU server.
• Faster, Predictable Answers: Fixed-format responses and two model calls per reply instead of three.
• Scheduled Crawler: A sitemap-driven crawler keeps the knowledge base in step with the website.

The impact

Impact