LearnVerse

The Challenge

Building a platform that could provide genuinely personalized education at scale, with AI that understands each student learning style, pace, and knowledge gaps, required solving problems at the intersection of AI, education science, and user experience. The AI tutoring system needed to be pedagogically sound — not just providing correct answers, but guiding students through the reasoning process in a way that builds genuine understanding. This is fundamentally different from a chatbot; the AI needed to know when to give hints, when to explain a concept differently, and when to let the student struggle productively. The adaptive learning engine needed to build accurate models of each student knowledge state from relatively few interactions, since students have limited patience for "diagnostic" assessments. The platform needed to work reliably in classroom settings where 30+ students might be using it simultaneously on school-provided Chromebooks with limited processing power and inconsistent WiFi. Accessibility was non-negotiable — the platform needed to meet WCAG 2.1 AA standards and support students with learning disabilities including dyslexia and ADHD through customizable interfaces and alternative content presentations. Teacher adoption was a critical concern; many teachers are overwhelmed by technology tools, so the interface needed to provide powerful insights with minimal setup and learning curve. Finally, student data privacy under Quebec Law 25 and COPPA (for users under 13) imposed strict requirements on data handling, AI training, and parental consent workflows.

Our Solution

We developed a custom adaptive learning engine built on a Bayesian Knowledge Tracing (BKT) model enhanced with deep learning, which maintains a probabilistic model of each student knowledge state across hundreds of granular learning objectives. The system updates its model with every student interaction — correct answers, incorrect answers, time spent, hints requested, and even mouse hover patterns that indicate uncertainty. The AI tutoring system uses GPT-4 with a carefully engineered system prompt and conversation management layer that implements the Socratic method: asking guiding questions rather than providing direct answers, offering progressively more specific hints when students are stuck, and providing encouraging feedback calibrated to the student emotional state (detected through response patterns and language). We built a "pedagogical guard" layer that reviews every AI response against educational best practices before displaying it to students, catching any instances where the AI might accidentally reveal an answer or use age-inappropriate language. The platform was built on Next.js with a progressive web app architecture that works on low-spec Chromebooks and can function partially offline (crucial for schools with unreliable internet). Real-time collaboration uses WebRTC for peer-to-peer whiteboard sharing with a Supabase backend for state persistence. The teacher dashboard provides at-a-glance class health metrics with drill-down capabilities, requiring zero configuration — teachers simply create a class code and students join. The accessibility system includes a dyslexia-friendly font option, customizable color themes, text-to-speech integration, and reduced-motion modes. Student data is stored in Supabase hosted in Canada, with automated data retention policies, parental consent management, and complete data export/deletion capabilities. The development took 10 months with a team of 5 developers and 1 UX researcher who conducted bi-weekly testing sessions with actual teachers and students throughout the build.

Results

Tech Stack