Project Overview and Business Context
The client was an edtech startup targeting GCSE and A-level students in the UK. Their existing product was a content library: video lessons and practice questions organised by topic, with no adaptation to individual performance. Learner engagement dropped sharply after the first two weeks as students either found content too easy (high-performing students) or became stuck and disengaged (students with specific knowledge gaps). The brief was to build an adaptive layer that monitors performance at the concept level, identifies specific knowledge gaps, and routes each learner to the content and practice that addresses their particular needs at the appropriate difficulty level. The system needed to do this without requiring a teacher to manually set learning paths, making it viable as a self-directed revision tool. UK GDPR obligations for student data are specific: students under 13 require parental consent for data processing, and data collected for educational purposes must not be used for commercial profiling. The system uses PostHog for product analytics (event-level, no personal identifiers) and Supabase for learning records, with data processing documented and parental consent flows built for the under-13 use case. GDPR-compliant data handling is particularly important in education, where schools and academy trusts applying the ICO's Age Appropriate Design Code have elevated data protection standards.
Technical Architecture and Stack Decisions
The application is built on Next.js deployed on Vercel, with Supabase handling authentication, learning record storage, and content metadata. The content library is stored in Supabase with structured metadata per item: subject, topic, concept, difficulty level (1-5), content type (video, text, worked example, practice question), and prerequisite concept tags. The adaptive engine is a server-side Next.js API route that maintains a learner knowledge model: a per-concept proficiency estimate (0-1 scale) updated after every practice question based on correctness, response time, and hint usage. The knowledge model is stored in Supabase and updated in real time. Content selection uses a Bayesian knowledge tracing approach to estimate concept mastery, then applies a selection algorithm that prioritises: concepts with the largest gap between estimated mastery and target mastery, concepts that are prerequisites for unmastered downstream concepts, and content types that have historically performed well for this learner (calculated from response time and correctness patterns). GPT-4o is used for dynamic explanation generation: when a learner encounters a concept for the first time or repeatedly answers incorrectly, the system generates a personalised explanation calibrated to their demonstrated vocabulary level (inferred from prior interaction patterns) and using analogies drawn from their subject preferences recorded during onboarding. PostHog tracks funnel metrics, session length, and content completion rates at the aggregate level, with no personal identifiers in the analytics stream.
Key AI and ML Components
GPT-4o handles adaptive explanation generation, which is the highest-value AI component for learner outcomes. The system prompt instructs the model to explain the target concept at a specified vocabulary level (rated 1-5 based on the learner's demonstrated reading level from prior interactions), to use an analogy from a specified domain (science, sport, music, technology, based on learner preferences), and to produce the explanation in under 150 words with a concrete worked example. The explanation is generated server-side and cached per concept per vocabulary level, so repeat exposures to the same concept do not incur repeated API calls. The Bayesian knowledge tracing model is a statistical component, not an LLM: it updates concept mastery estimates using correctness signals and a configured slip and guess parameter per question. This model is computationally cheap, runs entirely in the Next.js API layer without external API calls, and produces reliable mastery estimates that improve with each practice attempt. GPT-4o is also used for question generation: given a target concept and difficulty level, generate three novel practice questions with correct answers and common wrong-answer distractors. Generated questions are reviewed by a subject specialist before being added to the live question bank, maintaining content quality without manual authoring at scale.
Challenges Solved and How
Cold start is a classic problem in adaptive learning: the system has no data about a new learner, so initial content selection is uninformed. The build addresses this with a structured 15-minute diagnostic assessment covering the top-level concept hierarchy for the subject, using performance on diagnostic questions to seed the initial knowledge model before the adaptive engine takes over. The diagnostic is presented as an engaging challenge rather than a test, reducing the anxiety that suppresses accurate performance signals. Learner motivation varies significantly with content difficulty. Content that is too easy produces boredom and disengagement; content that is too hard produces anxiety and drop-off. The adaptive engine targets the zone of proximal development: content at the difficulty level just above the learner's current demonstrated mastery. PostHog data from early testing showed optimal engagement at a success rate of 70-75% on practice questions, informing the difficulty targeting algorithm. GDPR compliance for under-13 users required a separate consent flow for parental consent before account creation, a restricted data processing mode that excludes learning records from any analytics processing, and a parental dashboard showing learning activity data in summary form without individual question responses.
Outcome and Measurable Results
The client ran a 10-week pilot with 340 GCSE students across four schools. Average session length increased from 12 minutes (non-adaptive version) to 28 minutes (adaptive version), indicating significantly higher engagement. Practice question accuracy improved by an average of 23 percentage points over the pilot period for learners who completed at least 5 hours of adaptive practice, compared to 11 percentage points for a control group using the non-adaptive content library. Subject teacher feedback indicated that learners using the adaptive system arrived at classroom sessions with more targeted questions about specific sticking points, suggesting the system was accurately identifying and communicating knowledge gaps. PostHog data showed a 7-day retention rate of 61%, significantly above the 34% baseline of the non-adaptive product.
Lessons for Similar Projects
The knowledge model is more important than the content. An adaptive system is only as good as its ability to accurately estimate what a learner knows. Invest in the knowledge model design, the diagnostic assessment, and the question calibration before worrying about content volume. Content quality matters more than content quantity for adaptive systems. A library of 200 high-quality, accurately difficulty-rated items performs better than a library of 2,000 uncalibrated items. Plan for a content quality sprint alongside the technical build. Build the feedback loop for question quality early. Flagging poor questions, tracking question difficulty empirically from learner performance, and removing or revising underperforming questions is an ongoing operation that needs a process and a UI from launch day. GDPR compliance for educational products is non-trivial. Engage with the ICO's Age Appropriate Design Code early and build parental consent flows as a first-class feature, not an afterthought.