Traditional online learning platforms often treat every student the same way.
Two students may receive identical lessons, exercises, and recommendations despite having completely different strengths, weaknesses, and learning histories.
I wanted to explore how machine learning and educational analytics could be combined to create a more personalized learning experience.
This led to the development of Elo Learn, an adaptive learning prototype built using FastAPI and Streamlit that models student mastery, recommends personalized learning paths, explains its decisions, and schedules review sessions automatically.
The goal was not simply content recommendation, but building an end-to-end educational intelligence system capable of understanding how students learn over time.
Most learning platforms rely on simple completion metrics:
Completed a lesson
Passed a quiz
Watched a video
These signals provide limited insight into actual understanding.
A more effective system should answer questions such as:
What concepts has a student truly mastered?
Which prerequisite topics are missing?
What should they learn next?
Why was a recommendation made?
When should concepts be reviewed?
Building a system capable of answering these questions became the core motivation behind Elo Learn.
The platform combines several AI-driven components into a unified pipeline.
Student Interactions
│
▼
Bayesian Knowledge Tracing
│
▼
Student Mastery Estimation
│
┌───────────┼────────────────────┐
│ │ │
▼ ▼ ▼
Embeddings Knowledge Graph Review Engine
│ │ │
▼ ▼ ▼
Recommendations Readiness Spaced Repetition
│
▼
Explainable Learning Path
Instead of relying on a single metric, Elo Learn combines multiple educational signals to generate recommendations.
The first challenge was estimating student mastery.
To accomplish this, I implemented Bayesian Knowledge Tracing (BKT).
BKT models the probability that a student has mastered a concept based on their historical interactions.
Rather than treating learning as binary, the model continuously updates confidence scores as new evidence becomes available.
This allows the platform to distinguish between:
Concepts that are mastered
Concepts requiring reinforcement
Concepts that have not yet been learned
Knowledge tracing provides mastery estimates, but it does not capture broader behavioral similarities.
To address this, I created student and topic embeddings.
These representations allow the system to:
Compare learning patterns between students
Identify similar learners
Discover related concepts
Improve recommendation quality
Embedding-based representations help move beyond simple score tracking and enable richer personalization.
Building a Knowledge Graph
One of the most important components of Elo Learn is its concept graph.
Learning rarely happens in isolation.
For example:
Algebra
↓
Functions
↓
Calculus
A student struggling with Calculus may actually have gaps in prerequisite concepts.
The knowledge graph enables the platform to:
Identify dependencies between concepts
Calculate readiness scores
Recommend prerequisite material
Generate remediation pathways
This makes recommendations significantly more educationally meaningful.
Recommendation systems often operate as black boxes.
I wanted Elo Learn to provide transparency.
Instead of simply recommending a topic, the platform explains:
Why the topic was selected
Which concepts influenced the recommendation
Current mastery levels
Confidence scores
Prerequisite readiness
This creates trust and helps both students and instructors understand the reasoning behind recommendations.
Learning does not end after a concept is mastered.
Without reinforcement, knowledge decays over time.
To address this challenge, Elo Learn incorporates an SM2-inspired spaced repetition algorithm.
The review engine:
Tracks concept mastery
Calculates optimal review intervals
Identifies overdue concepts
Prioritizes revision sessions
This allows students to retain knowledge more effectively while reducing unnecessary review effort.
The platform was designed not only for students but also for educators.
Instructor dashboards provide:
Cohort Analytics
Track overall class performance and mastery trends.
At-Risk Student Detection
Identify students requiring intervention.
Weak Topic Identification
Highlight concepts causing widespread difficulty.
These analytics transform raw student interactions into actionable educational insights.
Backend Architecture
The backend was built using FastAPI and organized into modular components.
backend/
frontend/
recommendation_engine/
knowledge_graph/
ml_models/
datasets/
tests/
The system exposes endpoints for:
Mastery estimation
Recommendation generation
Knowledge graph reasoning
Cohort analytics
Review scheduling
This separation makes experimentation and future extensions easier.
The final system includes:
Bayesian Knowledge Tracing
Student and concept embeddings
Knowledge graph reasoning
Explainable recommendations
Spaced repetition scheduling
Cohort analytics
Instructor dashboards
FastAPI REST APIs
Streamlit visualization interface
Building Elo Learn revealed several interesting challenges.
Educational Data Is Complex
Student performance is not always a reliable indicator of understanding.
Explainability Is Essential
Recommendations become significantly more useful when accompanied by reasoning.
Knowledge Graph Design Matters
Incorrect prerequisite relationships can produce poor recommendations.
Multiple Signals Outperform Single Metrics
Combining mastery estimates, graph reasoning, and embeddings produced more meaningful recommendations than any individual component.
Several directions could make Elo Learn even stronger:
Graph Neural Networks for concept reasoning
Deep Knowledge Tracing using Transformers
Real-time adaptive assessments
Reinforcement learning for recommendation policies
Large-scale deployment with cloud infrastructure
Integration with Learning Management Systems
Elo Learn began as an exploration of adaptive learning systems and evolved into a complete educational intelligence platform.
Check it out here -> https://github.com/Shridipa/Elo-Learn
By combining Bayesian Knowledge Tracing, knowledge graphs, recommendation systems, explainability, and spaced repetition, the project demonstrates how multiple AI techniques can work together to support personalized education.
The biggest lesson from building Elo Learn was that effective learning recommendations require more than predicting what a student might like. They require understanding what a student knows, what they are ready to learn next, and why a particular learning path makes sense.