Misconception classifier
Every wrong answer gets classified into one of 200+ known misconceptions, not just 'wrong.' The taxonomy was built with the learning team from a decade of student work.
Loading…
Work/EdTech · 2M learners
EdTech · 2M learners · 2024-2025
A K-12 math platform was failing 40% of its learners with a one-size-fits-all curriculum. We built an adaptive AI tutor that calibrates difficulty per learner and explains stuck moments.
A K-12 math platform serving 2M learners had a one-size-fits-all curriculum. 40% of learners stalled at one of three predictable choke points (fractions, algebraic substitution, geometric reasoning) and churned within 30 days. The product team had been adding more practice problems for two years. Nothing moved the metric. The Head of Learning had a sharper question: what if difficulty calibrated to the learner instead of the grade?
We built an adaptive engine that tracks where each learner gets stuck, generates explanations targeted to the specific misconception (not the topic), and adjusts difficulty per-problem rather than per-grade. The model retrieves prior worked examples from the curriculum corpus, finds the closest analog to the learner's mistake, and walks them through it in their own grade-appropriate language. Two-thirds of churn came from three choke points. Targeting those alone delivered most of the gain.
Every wrong answer gets classified into one of 200+ known misconceptions, not just 'wrong.' The taxonomy was built with the learning team from a decade of student work.
Claude pulls a worked example from the curriculum corpus that matches the misconception and rewrites it in the learner's reading level. Same math, right level.
The next problem is selected to be solvable by the learner's current capability, not the grade-level average. Stuck learners get unstuck instead of buried.
Visible progress, where the learner is stuck this week, and what got unstuck. Parents see the work.
Built with the learning team. Curriculum corpus indexed and tagged.
Limited rollout to one grade band. Eval pipeline weekly from day one.
All K-8 math grades. Eval and content tuning.
Dashboards. Scaled to 2M learners. Retainer for ongoing eval.
Across all grade levels. Largest gains on previously stuck learners.
Stuck-then-churn cohort cut by more than half. $9M+ retained revenue.
Best predictor of long-term retention.
We've had AI features for two years that nobody used. This is the first thing the team built that learners and parents both notice.