At LMDDC, we have been building [AI]lice for more than two years and running it in production since December 2025. It is an AGPL-licensed RAG application that lets any educator build a chatbot grounded in their own course material, hosted entirely on LMDDC’s infrastructure. Teachers upload documents or connect a Moodle course, write the system instructions that define how the chatbot behaves pedagogically, and publish it to their learners. Every answer cites the material it came from, and no conversation is stored by default.
This talk is about what it takes to build and run that in-house. We present the architecture and the open components it relies on, the constraints that sovereignty and teacher oversight imposed on the design, and what self-hosting actually demands in hardware, energy and maintenance. We close with where we intend to take the project next, and the questions we have not yet answered.
Attendees will leave with a concrete reference point for sovereign AI in education: what the stack is made of, what it costs to operate, which trade-offs cannot be avoided, and where the project is heading.