Personalizing Learning for 50,000 Students

The Problem

Summit's online platform served 50,000 students but offered a one-size-fits-all experience. Completion rates sat at 34%, and students frequently dropped off at difficulty spikes. The platform needed personalization at scale without exploding costs.

Our Approach

Designed a credit-efficient personalization engine. Rather than running full LLM calls for every interaction, we built a tiered system: lightweight models handle routine interactions, with LLM calls reserved for complex explanations and adaptive content generation. Credits are budgeted per-student with automatic optimization.

The Solution

Implemented a three-tier tutoring system: (1) Rule-based responses for common questions, (2) Small model for personalized hints and encouragement, (3) Full LLM for complex explanations, concept mapping, and adaptive content. A credit optimizer dynamically shifts students between tiers based on need and budget.

Results

• Course completion rates improved from 34% to 58%
• Student satisfaction scores up 45%
• Per-student AI cost maintained under $3/month
• Platform retention improved by 62%