OpenVidya grounds tutoring in CBSE curricula

Hot take: the architecture is sensible, but the product is still too broad and too early to win on “India-first AI tutor” alone.

• –Architecture: the three-layer split across curriculum grounding, pedagogical mode, and runtime narration is the right shape, but it looks over-modeled for an early release; I would prove one end-to-end lesson loop before expanding the mode matrix further.
• –Product focus: start with teachers or coaching centers, not students; they are the clearest buyers, they can judge whether the content is board-aligned, and they can give fast feedback on lesson quality and question quality.
• –Evaluation: measure factual grounding error rate, chapter coverage, question validity, teacher-rated lesson quality, pre/post test lift, one-week retention, and time saved to prepare a lesson compared with a generic tutor.
• –Dataset: the highest-value open inputs are NCERT textbooks, NCERT exemplar problems, CBSE sample papers, NCERT lab manuals, NIOS materials, and any publicly available board-style question banks with clear licensing; add a schema for source provenance so the KB is auditable.
• –README/demo: the repo needs a concrete demo walkthrough, supported grades/subjects, example outputs for each mode, and a clearer “what is shipped vs planned” section; right now the promise is stronger than the evidence.
• –README/demo: there is also a license mismatch to fix, since the README says MIT while the repo page shows AGPL-3.0, which will confuse contributors and adopters.
• –Positioning: the strongest claim is not “AI tutor for India,” but “curriculum-constrained lesson generation for CBSE/NCERT workflows.”