Back to Projects
🎓

Elo Learn

AI-Powered Adaptive Learning Platform with Knowledge Tracing and Personalized Recommendations

Status

Production Ready

Timeline

5 months

Role

Full Stack ML Engineer

Problem Statement

Traditional e-learning platforms fail to adapt to individual learners:

  • One-size-fits-all curriculum doesn't account for learner skill levels and knowledge gaps
  • No intelligent sequencing — students struggle because prerequisites aren't taught first
  • Poor engagement from high dropout rates due to mismatched difficulty
  • Inefficient learning — students waste time on content they already know

Target Users: Online education platforms, corporate training, personalized tutoring systems.

Research Motivation

Why Knowledge Tracing?

Knowledge tracing models estimate student knowledge states from interaction history. By modeling "skill mastery," we can predict which concepts students will struggle with and adapt content in real-time.

Why Recommendation?

Recommender systems sequence learning content by difficulty and prerequisite requirements. Combining knowledge tracing with collaborative filtering personalizes learning paths for maximum engagement and retention.

System Architecture

Knowledge Tracing Pipeline

Interaction Logging: Capture question attempts, correctness, response time per skill

Skill Graph: Model prerequisite relationships between concepts (DAG structure)

Knowledge State: Estimate mastery probability for each skill using Bayesian inference

Prediction: Forecast success probability on new questions given current knowledge state

Recommendation Engine

Content Sequencing: Topologically sort skill graph + difficulty levels

Difficulty Balancing: Apply spaced repetition — revisit mastered skills with increasing intervals

Collaborative Filtering: Recommend content based on similar learners' paths (matrix factorization)

Cold-Start Handling: Use skill prerequisites when learner history is limited

Technical Implementation

Backend Stack

  • Framework: FastAPI
  • Knowledge Tracing: PyBKT (Bayesian Knowledge Tracing)
  • Recommendations: Surprise library, matrix factorization
  • Database: PostgreSQL
  • Graph DB: Neo4j (skill prerequisites)
  • Caching: Redis for real-time learner states

Frontend Stack

  • Framework: Streamlit (interactive UI)
  • Visualization: Plotly for progress tracking
  • Skill Tree Viz: Cytoscape.js for dependency graphs
  • Quiz Interface: Custom React components
  • Mobile: React Native companion app
  • Deployment: Docker + Kubernetes

Technical Challenges & Solutions

Challenge 1: Cold-Start Problem

Problem: New learners have no interaction history, making knowledge tracing unreliable.

Solution: Implemented hybrid cold-start: use diagnostic quiz to initialize knowledge state, then refine with learner interactions. Diagnostic scores correlate (r=0.78) with predicted mastery.

Challenge 2: Spaced Repetition at Scale

Problem: Computing optimal spaced repetition schedules for thousands of learners and concepts is computationally expensive.

Solution: Pre-computed spaced repetition intervals using forgetting curve parameters; Redis caching ensures O(1) lookups for personalized schedules.

Challenge 3: Fairness & Engagement

Problem: Overly challenging content discourages learners; overly easy content fails to engage.

Solution: Implemented difficulty calibration — target 70% success rate per learner. A/B tested and confirmed 15% higher completion rates vs. fixed-difficulty control.

Methodology

Evaluation Framework

Evaluated across learning science metrics:

  • Knowledge Tracing Accuracy: AUC-ROC for predicting next question correctness
  • Recommendation Quality: NDCG@5 for content sequencing; A/B testing on engagement
  • Learning Outcomes: Post-test mastery, retention after 2 weeks
  • User Engagement: Completion rates, session duration, dropout rates

Study Design

Pilot with 500+ learners across 10 courses. Randomized controlled trial: adaptive vs. fixed curriculum. Measured learning gains and engagement metrics.

Results & Impact

Knowledge Prediction AUC

0.82

Strong predictive signal

Engagement Improvement

+23%

Completion vs. control

Learning Gains

+18%

Post-test improvement

Key Findings

  • • Adaptive difficulty increases engagement by 23% and completion by 15% vs. fixed curriculum
  • • Knowledge tracing accurately predicts question difficulty (AUC=0.82); enables real-time adaptation
  • • Spaced repetition + difficulty balancing improves retention by 31% at 2-week follow-up
  • • Learners value transparent skill mastery progress tracking (NPS +42)

Future Work

Deep reinforcement learning for optimal content sequencing (Thompson sampling)

Social learning: peer collaboration and knowledge sharing recommendations

Multi-domain transfer learning: leverage mastery in one subject to accelerate another

Neuroplasticity-informed adaptation: personalize learning style (visual, kinesthetic, etc.)

Integration with existing LMS platforms (Canvas, Moodle, Blackboard)

Predictive dropout prevention with early intervention strategies

Links & Resources

Built with Next.js, TypeScript, and TailwindCSS.

© 2026 Shridipa Dhar