A university-built AI tutor just proved that the biggest language model isn't always the one that makes the cut.
Researchers built CourseChat, an on-premises retrieval-augmented generation tutor for six undergraduate business course sections, each tracked by its own course reference number. The system runs on twin-edge AI hosts behind a campus web gateway, pairing a FastAPI service with a local vector database and a local large language model served through Ollama, with plans to embed it in Moodle. The team ran two rounds of model bake-offs plus a separate test of how faithfully different setups stuck to the source evidence, then audited real conversations and quizzes. Several larger models were too slow for classroom use; a 12B model and a 7B model passed the speed gate, and a mixture-of-experts model fixed some errors while introducing new ones, so the team kept its existing 8B production model instead of swapping it in.
This is a case study in the unglamorous reality of deploying AI tools under real constraints. A campus can't ship a chatbot that takes ten seconds to answer a syllabus question, and it can't send student conversations to an outside vendor's cloud. The paper treats model size, evidence retrieval, and serving software as one combined decision rather than picking a flashy model and hoping the infrastructure catches up. That's a more honest framing than most splashy model announcements lately.
The researchers are careful to note what they haven't shown: whether students actually learn more, how the system holds up under a full class's peak traffic, or whether it's ready for a wider campus rollout. That's the kind of detail that's easy to skip when writing a press release.