An education technology platform serving 500,000 K-12 students needed to move beyond one-size-fits-all content delivery. Students at different skill levels, learning speeds, and engagement patterns were receiving the same content sequence. Struggling students fell further behind while advanced students were bored. The platform deployed AI-driven personalization to adapt the learning experience to each student.

The Architecture

The system maintains a student knowledge model, selects appropriate content, and generates personalized practice materials.

Knowledge model - For each student, the system maintains a probabilistic model of mastery for each skill in the curriculum graph (a directed graph where skills have prerequisite relationships). As students answer questions, the model updates mastery estimates using Bayesian knowledge tracing. A student who answers three consecutive questions correctly on a skill has their mastery estimate increased; incorrect answers decrease it. The model also infers mastery of prerequisite skills from performance on dependent skills.

Content selection engine - Based on the student’s current mastery state, the system selects the next learning activity from the content library. The selection algorithm targets the “zone of proximal development” - skills where the student has mastered prerequisites but has not yet mastered the skill itself. Content difficulty is matched to the student’s estimated ability level, targeting a 70-80% success rate that maintains engagement without causing frustration.

Practice generation - An LLM generates personalized practice problems calibrated to the student’s level. For a student working on two-digit multiplication who struggles with carrying, the system generates problems that specifically practice carrying. For a student reading at grade level, the system generates comprehension questions about topics the student has shown interest in.

Engagement monitoring - The system tracks engagement signals: time-on-task, response latency, hint usage, and session frequency. Declining engagement triggers adaptive responses: reduce difficulty temporarily, switch content modality (video instead of text), introduce gamification elements, or suggest a break.

Key Lessons

The prerequisite graph was the foundation. Without accurate prerequisite relationships between skills, the system could not determine appropriate next steps. Building and validating the curriculum graph required extensive collaboration with curriculum designers and took four months. Errors in the graph (missing prerequisites, incorrect prerequisite chains) caused students to encounter content they were not prepared for.

The 70-80% success rate target was critical for engagement. Students below 50% success rate showed signs of frustration (longer response times, more hint requests, shorter sessions). Students above 90% success rate showed signs of boredom (faster but more careless responses, shorter sessions). The 70-80% range maintained optimal engagement and learning velocity.

LLM-generated practice problems required careful validation. The system could generate unlimited practice problems, but some contained errors (incorrect answer keys), ambiguous wording, or inappropriate difficulty. A quality pipeline validated each generated problem: checking answer correctness, estimating difficulty, and filtering for age-appropriate content. Approximately 15% of generated problems were filtered out.

Teacher dashboards drove adoption. Teachers initially resisted the system because they felt replaced. Building dashboards that showed each student’s progress, identified struggling students, and suggested intervention points made teachers feel empowered rather than replaced. Teacher adoption jumped from 40% to 85% after the dashboard launch.

Results

Student learning gains (measured by pre/post assessment) improved by 23% compared to the non-personalized experience. Student engagement (average sessions per week) increased by 35%. The percentage of students completing course objectives increased from 62% to 78%. Teacher satisfaction scores improved from 3.1/5 to 4.2/5 after dashboard deployment. The platform’s student retention rate improved by 18%.