personal AI tutor — multimodal input + learner model + safety
multimodal input
text (web chat)
voice (LiveKit + Whisper)
photo math (dots.ocr)
tutor policy (LangGraph)
read signal (correct / partial / stuck)
select next concept (graph walk)
scaffold / hint / celebrate
learner model
Bayesian knowledge tracing
per-concept mastery
spaced-repetition (FSRS)
update after each interaction
weekly progress report
parental dashboard
curriculum graph
Neo4j of 50-150 concepts
prerequisite edges
OER content attached
K-12 algebra OR intro Python
pick one for depth
walk algorithm:
pick lowest-mastery concept
whose prereqs are mastered
OER sources:
OpenStax, Open Textbook,
Khan Academy lite
safety + memory
Llama Guard 4 + age filter
COPPA-aware retention (1y)
memory: episodic + semantic
parent-accessible deletion
learner-scoped isolation
efficacy study (2 weeks, 10 learners):
pre-test + 6 sessions + post-test
adaptive cohort vs linear baseline
Khanmigo / Duolingo Max / LearnLM shape