Case Study · Higher Education

A Closed-RAG AI Tutor Inside a University's Moodle

A Central American university wanted students to get AI help with coursework, without the risk of an open model inventing answers about their own syllabus. Closed RAG was the answer.

By Brandon Da Costa, Founder, Digiton Dynamics

The risk with an open model in education

Ask a general-purpose AI model a question about a specific course and it will answer confidently whether or not it actually knows the syllabus, the grading rubric, or the professor's own version of the material. In a university setting that is not a minor annoyance, it actively misleads students at the exact moment they are trying to learn something correctly. The university's concern was not 'should students use AI,' it was 'how do we stop AI from confidently teaching them the wrong thing.'

Why closed RAG, specifically

Closed RAG means the assistant only answers from a defined, approved set of course material, indexed inside the retrieval layer, rather than from the model's open training data. If a question falls outside what the course documentation covers, the system says so instead of guessing. That constraint is deliberate and, for education, non-negotiable: a system that occasionally invents a plausible-sounding but wrong answer is worse than a system that sometimes says 'that's not covered in this course,' because the second failure mode is honest and the first is not.

Building it inside Moodle

The university runs on Moodle, so the assistant had to live where students already were rather than asking them to adopt a separate tool. I built the integration to sit inside that existing environment, delivered through a partner on the ground in Latin America, so the rollout matched how the university already operates rather than forcing a new workflow onto students and faculty mid-semester.

What 'delivered through a partner' means in practice

I want to be specific about the delivery model here, since it differs from a project where Digiton owns the entire client relationship end to end. This assistant was built and delivered in partnership with a local organization already working with the university, which handled parts of the on-the-ground relationship and rollout while I focused on the architecture and the technical build. That is a common pattern for international education work: the technical build benefits from being paired with a partner who understands the institution's actual operating context, calendar and constraints, rather than me guessing at them from a distance.

Why this pattern fits universities specifically

Universities are an unusually good fit for closed RAG because the source material is already curated: a syllabus, a set of readings, assignment instructions, lecture notes. Someone has already done the work of deciding what is authoritative for a given course, which is exactly the kind of bounded, high-quality source set that closed RAG needs to be reliable. Compare that to a domain where the 'right' answer is genuinely ambiguous or constantly shifting, and the same architecture would be a much harder sell. Education, done this way, is close to the ideal use case for the pattern.

What students actually get

A student can ask a question about course material in plain language and get an answer grounded specifically in that course's approved content: lecture material, readings, assignment instructions. The assistant is explicit about its boundary. It is not trying to be a general tutor for every subject in the world, it is trying to be a reliable, always-available reference for one course's actual material, which is a narrower and much more honest promise than most AI-in-education pitches make.

Where it stands

The assistant is deployed inside the university's Moodle environment. I am deliberately not making satisfaction claims here, because quality tuning on real student usage is an ongoing process, and the honest status is 'deployed,' not 'solved.' Education is one of the domains where I would rather under-claim than over-claim, since the cost of a wrong claim compounds against real students.

The honest state of quality tuning

Closed RAG reduces the risk of confidently wrong answers, but it does not eliminate the harder work of tuning retrieval and response quality against how real students actually ask questions, which rarely matches how course material is written. That tuning is iterative: watching real usage, identifying where the assistant retrieves the wrong passage or answers too narrowly, and adjusting the retrieval and prompting accordingly. As of this writing that process is ongoing, which is exactly why I am careful to say 'deployed' rather than claim a finished, optimized outcome. I would rather under-promise here than round up.

The broader pattern

Closed RAG is the right architecture anywhere an organization needs AI to be helpful without ever being confidently wrong about material it should know cold: internal training, compliance material, onboarding, or any domain where 'I don't know' is a better answer than a guess. If your organization has a body of material students, staff or customers need answered accurately rather than plausibly, that is a system worth building properly. Reach out if that is the problem you are facing.

Frequently asked questions

Related