AI Trailblazer
Piloting a closed-loop AI tutor inside online modules, with custom JavaScript that keeps the model answering from verified module content rather than freestyling.
The brief
The AI Trailblazer modules were the university's test of an integrated LLM inside the LMS, trained on the content within each learning room so the model could only draw on the material students had actually been taught. The idea was to give learners a tutor for general queries. The real design question sat elsewhere: what could this actually do for learning if you built the prompts properly?
My role
I did not build the underlying tool. I designed the pedagogical framework around it and built the custom JavaScript that injects structured prompts, so the model only responds within the content it has been given.
How I approached it
I focused on three things a well-scoped AI tutor can actually do for a learner: reflection, summarisation, and short branching scenarios. For each I wrote a fixed prompt that framed the model's role, and used a custom textbox so students could type the specific week or topic they were working on and press a button to send the full prompt, their input already sitting inside it.
I built the same pattern for reviewing a student's written reflection against established reflection models (Kolb, Gibbs), so the AI was comparing the student's own reflection against a taught framework rather than generating one from scratch. The design work was in the fixed wording around the injection point, not the injection itself.
There is a working recreation of the two AI tools you can try, showing the prompt design and the injection flow. The prompts are lightly cleaned versions of what was actually deployed.
Outcome
The tools were built and ran through the pilot. The university subsequently chose not to take the wider AI tool forward institutionally, so the framework and injection tooling stand on their own rather than continuing in production. The prompt design work, and the pedagogical thinking around what an AI tutor should and should not be asked to do, is what I take forward from it.
What I would do differently
Given the same pilot again, I would push earlier on evaluating the model's actual answers against the fixed prompts, so we had evidence of where the scaffolding held and where it quietly slipped, rather than trusting the injection alone. That evaluation is more important than the injection itself.