#1Why generic models hit a ceiling in education
Education is contextually dense. A geometry tutor needs to know the sequencing of theorems in this particular curriculum, the misconceptions the student has shown this week, and the tone the school has approved. A generic model does none of those by default.
#2Curriculum as data, not just content
We treat the curriculum as a structured graph: skills, prerequisites, common misconceptions, exemplar solutions, and assessment items. That graph becomes the retrieval substrate for the model, ensuring every answer is grounded in the school's actual material.
#3RAG vs. fine-tuning vs. both
Retrieval-augmented generation gives the model fresh, curriculum-specific context at inference time. Fine-tuning gives the model the right voice, scaffolding behaviour, and refusal patterns. Most production tutors need both layered carefully — RAG for facts, fine-tuning for pedagogy.
#4Evaluating an AI tutor honestly
Evaluation is the hardest part of EdTech AI. Accuracy metrics miss the point — a tutor can be 'correct' and pedagogically harmful (by giving away answers). We pair automated evaluation with classroom A/B trials and qualitative review by teachers to assess whether the tutor actually moves learners forward.

