Back to Portfolio
HackathonCompleted

AI Learning Path Generator

Deerhack Hackathon

AIEducationPersonalizationMachine Learning
Developer & ML EngineerHackathon Project — 2022Nepal
AI Learning Path Generator

1st Place

Deerhack Hackathon

Real-time

Adaptive content delivery

ML-powered

Learning style classification

Personalized

Unique path per student

Overview

This project won 1st Place at the Deerhack Hackathon. We built an AI-powered personalized learning path generator that adapts to individual student learning styles based on quiz results and platform interactions. The system moves beyond one-size-fits-all curricula by dynamically adjusting content difficulty, format, and pacing based on each student's demonstrated comprehension patterns.

The education technology space is saturated with content platforms, but most treat all learners identically. Research consistently shows that students learn at different rates and through different modalities — some are visual learners, others learn through practice, and some need more repetition before concepts stick. Our system used machine learning to detect these patterns and respond accordingly.

The core innovation was a real-time adaptation engine that analyzed quiz responses not just for correctness but for response patterns (time taken, confidence levels, types of errors) to build a learning profile that improved with each interaction.

The Problem

Traditional online learning platforms deliver the same content in the same order to every student, regardless of their existing knowledge, learning speed, or preferred learning style. This leads to two failure modes: advanced students get bored with material they already understand, while struggling students fall behind without adequate support. Drop-off rates on MOOCs often exceed 90%, partly because the learning experience is not personalized to individual needs.

My Role

Developer & ML Engineer

I designed and built the ML pipeline for learning style classification and content recommendation. This included feature engineering from quiz interaction data, model training for learning style detection, and building the recommendation engine that mapped learning profiles to optimal content sequences.

The Approach

01

We built a three-layer system: data collection (capturing granular interaction data from quizzes and content consumption), learning profile construction (ML models that classify learning styles and knowledge gaps), and content recommendation (algorithms that select the next best content piece based on the student's profile).

02

The learning style classifier used features beyond simple right/wrong answers: response time per question, pattern of errors (conceptual vs. computational), self-reported confidence, and content format engagement (video watch time vs. text reading time). These features fed into a clustering algorithm that identified distinct learner archetypes.

03

Content recommendations used a hybrid approach combining collaborative filtering (what worked for similar learners) with knowledge graph traversal (prerequisite relationships between topics). This ensured recommendations were both personalized and pedagogically sound.

Key Features

What we built

Adaptive Quiz Engine

Quizzes that adjust difficulty in real-time based on response accuracy and patterns, maintaining optimal challenge level for each student.

Learning Style Detection

ML models that identify learning preferences from interaction patterns — visual, textual, practice-based, or mixed — and adapt content delivery accordingly.

Personalized Learning Paths

Dynamic curriculum sequencing that accounts for prerequisite knowledge, learning pace, and preferred content formats.

Progress Analytics

Dashboard showing students their learning trajectory, knowledge gaps, and recommended focus areas with clear visual indicators.

Tech Stack

Python
Scikit-learn
Flask
React
PostgreSQL
Collaborative Filtering
Knowledge Graphs

Key Lessons

What I took away from this project

The most valuable signals for personalization come from behavioral patterns, not self-reported preferences

Educational technology must balance personalization with pedagogical structure

Real-time adaptation requires lightweight models that can run inference quickly

Building a compelling demo for judges means showing the adaptation in action, not just explaining the algorithm

Want to build something similar?

I help companies scale their products and build high-performing teams. Let's discuss how I can help with your next project.

Get in Touch