Luminary — Adaptive EdTech Learning Platform

An online coding bootcamp was seeing 60% dropout before course completion. Their one-size-fits-all curriculum frustrated fast learners (bored) and overwhelmed slower ones (lost). Luminary replaced the linear syllabus with a knowledge-graph-driven adaptive engine that meets every learner exactly where they are — and moves at exactly the speed they can handle.

• 1

  A single linear curriculum cannot serve both a complete beginner and someone pivoting from a related field — the dropout cliff was real and measurable.

• 2

  Assessing true mastery (vs. guessing) requires more than MCQs — the system needed code-execution challenges evaluated for correctness and efficiency.

• 3

  Instructors were worried AI pacing would sideline their expertise — adoption depended on giving them visibility and control.

• 4

  Real-time collaborative exercises between learners in different timezones added significant architectural complexity.

We modelled the curriculum as a directed knowledge graph: each concept node has prerequisites and feeds into successors. A Bayesian Knowledge Tracing (BKT) model continuously estimates each learner's mastery probability for every node based on their exercise outcomes. The adaptive engine selects the next activity that maximises expected mastery gain given current confidence levels — avoiding both the 'too easy' boredom zone and the 'too hard' anxiety zone. Code challenges run in sandboxed Docker containers with AST-based evaluation; partial credit is awarded for correct logic with suboptimal complexity. Instructors see a live cohort dashboard showing mastery heatmaps and can inject 'mandatory' nodes for time-sensitive content.