The problem

Most AI tutors just answer questions or dump information — they don't verify you actually learned anything, and there's no persistent sense of progress beyond a checklist. I wanted a tutor that teaches in a genuinely Socratic, multi-turn way, tests comprehension before marking anything "done," and gives the learner something tangible to watch grow as proof of mastery — not just another progress bar.

Approach

Multi-turn Socratic AI tutor built on a TEACHING_STYLE protocol: real depth delivered in chunks, with a genuine comprehension check between each chunk before the tutor moves on.
A cognitive-science reinforcement loop layered on top of that — recall checkpoints, spaced repetition, and teach-back — so understanding is revisited, not just delivered once.
Syllabus-to-curriculum generation: users upload a syllabus (PDF/DOCX/TXT), Gemini parses it into subjects → topics → difficulty → prerequisites, each with its own AI-generated teacher persona.
Persistent 3D brain visualization in Three.js: proven topics render as glowing nodes, positioned by an AI-classified domain taxonomy (mechanism / conceptual / factual / skill / emotional) — a node only lights up after passing a test, never just from reading.
Voice-based mock interviews via Sarvam's voice models as a separate assessment surface, deliberately not wired into the core brain/learning loop.
Razorpay subscriptions with credit-based usage tiers (lib/plans.ts, lib/credits.ts) gate usage throughout, on top of Firebase auth.

Architecture

Diagram · loads on scroll

The keystone is that growth is earned, not given — the 3D brain only lights up a node after a topic passes its test, never just from being read or chunked through. Gemini turns an uploaded syllabus into a full curriculum tree with per-topic teacher personas; the Socratic tutor then delivers each topic in comprehension-gated chunks (the TEACHING_STYLE protocol) before a topic is even eligible for testing. A pass routes through an AI domain classifier that positions the new node in the brain by cognitive type — mechanism, conceptual, factual, skill, or emotional — so the visualization reads as an actual map of what's been mastered, not a progress bar. Voice-based mock interviews (Sarvam) sit alongside as a separate assessment surface, and Razorpay-backed credit tiers gate usage throughout.

Outcome

Live and monetised at claratto.com.
Real users proving topics through multi-turn Socratic sessions, not just marking lessons complete.
Working credit-based subscription billing via Razorpay, with Firebase auth end to end.

Lessons learned

Comprehension checks between chunks — not just at the end — is what makes a tutor feel Socratic instead of a lecture with a pop quiz bolted on.
Growth has to be provable, not just visual. Gating the brain's growth on passing a test (never on reading) is what keeps the visualization honest as a signal of mastery instead of decoration.
Splitting assessment (voice mock interviews) from the core learning loop kept both systems simpler than trying to unify them into one flow.