The Problem
Online learning has a completion problem. Most learners who start a professional upskilling course don't finish it. not because the material is too hard, but because the material is static. The same lesson is delivered to the overconfident beginner and the experienced practitioner who only needs targeted review. Neither gets what they need.
We set out to build an LMS that adapts to each learner in real time. adjusting content difficulty, surfacing targeted practice, and using AI to explain concepts in the learner's own context.
What We Built
The platform centres on an adaptive learning engine that continuously assesses learner performance and adjusts the path forward. Unlike traditional LMS systems that deliver course chapters in a fixed sequence, our engine computes a "knowledge gap score" per concept after each assessment and chooses the next module accordingly.
AI Practice Generation: rather than shipping static question banks, GPT-4o generates practice questions and case studies tuned to the learner's current level and their profession. A finance professional studying data analysis gets finance-domain examples; an engineer gets engineering contexts. Questions are generated fresh per session, preventing answer memorisation.
Conversational Coaching: each course module includes an AI tutor that learners can ask questions. The tutor has full context of the learner's performance history and can explain the same concept multiple ways, including "Explain this like I'm a complete beginner" and "Give me a challenging edge case." Powered by a streamed GPT-4o completion with RAG over the course material.
Cohort Learning: learners are matched into small cohorts (8-12 people) based on profession and experience level. Weekly cohort challenges create accountability without requiring synchronous meetings.
Employer Dashboard: for B2B sales, we built an employer-facing dashboard showing team learning velocity, assessment scores, and completion forecasts for managers.
My Role
I was responsible for the full technical architecture and hands-on engineering:
- Designed the adaptive learning algorithm (Bayesian knowledge tracing model, Python)
- Built the FastAPI backend. course management, assessment engine, progress tracking
- Implemented the streaming AI tutor with RAG pipeline (Supabase vector store + LangChain)
- Built the Next.js frontend from design to production
- Set up CI/CD, monitoring, and the Supabase database with RLS policies
- Led a team of two junior engineers through the build
Outcome
The platform launched with a cohort of 120 learners in a closed beta and reached 2,800 enrolled learners across 14 course categories within 9 months. Average course completion rate was 67%. well above the industry average of ~15% for self-paced online courses. NPS at 8-week mark: 72.
We ran a 12-month income comparison experiment: learners who completed courses on our platform reported an average salary increase of 22% at 6-month follow-up (self-reported via survey, n=380). We use this as our headline outcome metric in sales.