Open Category Submission

936 – Pebble

Pebble is an innovative AI-powered self-learning platform designed to transform how university students engage with educational content by addressing key challenges in higher education: lack of personalisation, ineffective revision practices, inconsistent engagement, and fragmented digital learning experiences.

The platform's core innovation lies in its integration of multiple learning modalities powered by advanced AI techniques. Users can upload their learning materials and gain access to the following features that aim to enhance their learning experience:

1. Multimodal Learning: Implements the VARK model (Visual, Auditory, Reading/Writing, Kinesthetic) by transforming educational content into interactive concept graphs for visual learners and AI-generated narrative podcasts for auditory learners, alongside traditional textual materials.
2. Adaptive Assessment Engine: Utilises a sophisticated "Planning First, Question Second" (PFQS) methodology combined with RAG (Retrieval-Augmented Generation) techniques to dynamically generate contextually relevant questions, adjust difficulty of subsequent questions based on performance, and provide personalised feedback.
3. Knowledge Tracking System: Employs a mastery-based progression model that monitors student understanding of granular learning objectives and adapts content accordingly.
4. Intelligent Chatbot Assistant: Features an AI-powered chatbot that leverages RAG to retrieve and synthesize relevant context from uploaded documents when answering student queries, with an optional web-search toggle that expands knowledge retrieval beyond course materials when enabled by the user.
5. Comprehensive Supporting Features: Integrates evidence-based learning tools including an integrated note-taking and flashcard system (leveraging active recall and spaced repetition), automatic cheatsheet generation, and a behavioral nudging notification system—all within a unified interface to eliminate the fragmented digital learning experience by providing a one-stop platform for all study activities.

Technically, Pebble leverages a microservice architecture for scalability, with specialised LLMs for different tasks (Gemini 2.0 Flash for concept mapping, GPT-4o/o3-mini for question generation and evaluation). Vector databases enable efficient semantic search capabilities for context retrieval.

The platform represents a significant advancement in AI-powered personalised self-learning by creating a cohesive learning ecosystem that adapts to individual learning preferences while promoting consistent engagement and effective knowledge retention through evidence-based educational principles.

• Pebble_Presentation.pdf

The development of Pebble was inspired by observing persistent challenges university students face despite technological advancements in education. Through literature review and observation, we identified four interconnected problems:
1. Lack of personalization: Traditional teaching methods adopt a one-size-fits-all approach that fails to accommodate diverse learning styles. Current educational platforms rarely integrate multiple modalities within a single ecosystem, leading to disengagement when instructional methods don't align with students' preferred learning styles (El-Sabagh, 2021).
2. Revision and retention challenges: The natural forgetting curve necessitates effective strategies to combat memory decay. Traditional methods emphasize memorization over understanding, resulting in surface-level comprehension. University students struggle to maintain consistent engagement across multiple courses, often resorting to cramming during examination periods (Gupta, 2024; Hennessy & Murphy, 2023).
3. Motivation barriers: Students' motivational values for academic subjects decrease over time. Passive revision techniques, lack of social interaction, absence of feedback, and unclear progress visibility lead to disengagement. Self-determination theory indicates that relatedness is a critical psychological need for motivation, which independent study often fails to address (Robinson et al., 2019; Bosch & Spinath, 2023).
4. Fragmented learning experience: Students constantly switch between disconnected platforms (note-taking tools, discussion forums, content portals), causing attention diversification and workflow inefficiencies that reduce productivity (Murray, 2021).

Our hypothesis is that an integrated platform addressing these specific challenges will significantly improve learning outcomes by:
1. Addressing personalization needs through multimodal content delivery (VARK model) with visual concept graphs and narrative podcasts alongside traditional text-based materials.
2. Overcoming retention challenges through evidence-based techniques like active recall, spaced repetition, and adaptive quizzing (PFQS methodology with RAG) that adjust to individual mastery levels.
3. Enhancing motivation through clear progress tracking with granular learning objectives, experiential points, achievement streaks, and targeted notification nudges based on behavioral science principles.
4. Eliminating fragmentation by unifying essential learning tools (content viewer, note-taking, chatbot, assessment) in a comprehensive ecosystem that minimizes context switching.

Our approach leverages evidence-based educational theories—VARK learning styles (Fleming & Mills, 1992), microlearning (Samala et al., 2023), active recall (Roediger & Karpicke, 2006), spaced repetition, goal-setting theory (McClelland et al., 1953), and behavioral nudging (Manuello, 2022)—while enhancing their implementation through AI technologies that make personalization and adaptation scalable and responsive to individual learning needs.

Technological Implementation
Pebble leverages a microservices architecture built with Python FastAPI, enabling modular development and scalability. Our solution integrates:
- Vector databases (Pinecone) for efficient semantic search and context retrieval
- Multiple specialized LLMs: GPT-4o/o3-mini for question evaluation and feedback, Gemini 2.0 Flash for concept mapping and podcast generation
- RAG-Fusion techniques for enhanced contextual retrieval
- Next.js/React frontend for responsive user experience
- MongoDB for user data and progress tracking
- AWS S3 for document and audio storage
- Docker containerization with Azure Container Apps for deployment
- Text to Speech APIs (elevenlabs) for engaging podcast generation
- Visualization libraries (D3.js) for concept graph generation)

Resources required include LLM API access, cloud infrastructure, and testing partnerships with educational institutions. We've validated market demand through literature review and preliminary user surveys.

Core Functionalities
- Interactive concept graphs for visual learning and knowledge organization
- Adaptive assessment system with personalized question generation and feedback
- AI-generated educational podcasts for auditory learning
- Integrated note-taking and flashcard creation
- Automated cheatsheet generation from course materials and notes
-AI chatbot for doubt clarification with document context
- Progress tracking with granular mastery assessment
- Notification system for engagement nudging

User experience is ensured through:
- Support for multiple learning preferences
- Clean, intuitive interface design
- Responsive performance through independent service scaling
- Comprehensive error handling

Effectiveness will be evaluated using:
- Learning metrics: Knowledge retention, mastery level progression, completion rates
- Engagement metrics: Time spent, return frequency, feature utilization
- AI performance metrics: RAGAS Faithfulness and RAGAS answer relevance scores for generated content
- User satisfaction: Comprehensive surveys measuring perceived usefulness and ease of use

Functional Architecture and Technical Workflow
Pebble employs a microservices architecture where distinct AI-powered services handle specific educational functions while communicating through well-defined APIs. The workflow begins with document processing and vectorization, followed by concept graph generation, which forms the scaffold for the adaptive learning experience. When users interact with learning objectives, the question service dynamically generates assessments, while the assessment service evaluates responses and updates mastery states.

Implementation of Innovative Features
The most innovative feature is the PFQS (Planning First, Question Second) methodology for adaptive question generation. This approach first generates a structured answer plan containing candidate answers, supporting evidence, and appropriate difficulty level based on user mastery. This plan then guides the generation of contextually relevant questions and evaluation criteria, resulting in more controlled and effective assessment.

Another key innovation is the hybrid RAG-Fusion search that enhances context retrieval by expanding queries and reranking results, ensuring more relevant context for both question generation and chatbot responses.

Function-Technology Relationship
1. Adaptive Question Generation
PFQS methodology with o3-mini for answer planning, Gemini 2.0 Flash for question generation
Accuracy improvement required

2. Contextual Retrieval
RAG-Fusion with Pinecone vector database and hybrid search
Fully implemented

3. Mastery Tracking
MongoDB with custom heuristic algorithm for mastery level calculation
Fully implemented

4. Concept Graph Visualization
D3.js for interactive graph rendering with topic extraction via Gemini 2.0 Flash
Accuracy improvement required

5. Podcast Generation
Gemini 2.0 Flash for narrative creation with ElevenLabs for voice synthesis
Fully implemented

6. Cheatsheet Generation
Gemini 2.0 with structured output formatting and PDF rendering
Fully implemented

7. Notification System
Azure Function App with Gmail API integration
Fully implemented

8. User Authentication
Clerk Auth with session management
Fully implemented

All components are deployed and operational on Azure Container Apps with independent scaling.

Pebble represents a significant innovation in educational technology by reimagining how AI can personalise the self-learning experience beyond simple content recommendation or question answering.

Innovative Approach to Learning Personalisation
Traditional adaptive learning platforms typically focus on content sequencing or quiz difficulty adjustment. Pebble takes a fundamentally different approach by personalising the entire learning modality to match individual preferences. By transforming the same educational content into visual concept graphs, narrative-driven podcasts, and interactive assessments, Pebble addresses the underlying cognitive diversity of learners rather than merely adapting content difficulty.

The PFQS (Planning First, Question Second) methodology represents a novel approach to question generation that fundamentally differs from conventional techniques. Instead of directly generating questions, our two-stage process first creates a structured answer plan with supporting evidence and difficulty parameters, which then guides precise question formulation. This ensures both contextual grounding and pedagogical alignment—a creative solution to the hallucination problems common in AI-generated educational content.

Creative Integration of Educational Theory and AI
Pebble creatively combines established educational principles with cutting-edge AI capabilities:
1. The microlearning approach implemented through interconnected concept nodes addresses cognitive load theory while leveraging graph visualization techniques
2. The mastery tracking system implements a nuanced heuristic that considers not just correctness but also question difficulty, attempt history, and learning progression
3. The notification system applies behavioral nudging principles through algorithmic timing of different types of engagement prompts

Perhaps most innovative is how Pebble's features form a cohesive learning ecosystem where outputs from one process become inputs to another—notes feed cheatsheet generation, topic extraction guides question generation, and performance tracking influences content difficulty—creating a virtuous cycle that addresses the fragmented nature of current educational technology.

This holistic approach represents a creative reimagining of how AI can transform learning from isolated technological interventions into an integrated, responsive educational environment.

Scalability Strategies
Pebble is architected with scalability as a foundational principle through several key strategies:
1. Microservices Architecture: Each functionality (Question Generation, Assessment, Retrieval, etc.) operates as an independent service that can be scaled based on specific demand patterns, preventing system-wide bottlenecks when particular features experience high usage.
2. Containerised Deployment: All services are containerised using Docker and deployed on Azure Container Apps with independent auto-scaling configurations (1-10 instances), enabling dynamic resource allocation as user demand fluctuates.
3. Database Partitioning: MongoDB is structured to allow for horizontal scaling through partitioning user data and track information, while Pinecone vector database handles efficient scaling of document embeddings.
4. Asynchronous Processing: Computationally intensive tasks like podcast generation and cheatsheet creation are handled asynchronously, preventing user experience degradation during peak loads.

Potential bottlenecks are addressed through:
- Caching frequently accessed content and embeddings
- Implementing rate limiting for intensive operations
- Optimising LLM prompts for token efficiency to reduce API costs at scale
- Employing hybrid search techniques that balance performance and accuracy

For sustainability and long-term engagement:

Environmental sustainability:
- Optimise compute resources through efficient prompt engineering to reduce unnecessary LLM calls
- Implement resource-aware scaling policies to balance performance with energy consumption

User engagement sustainability:
- Expand the notification system to include personalised learning recommendations based on user progress patterns
- Implement a continuous feedback loop that refines the adaptive difficulty algorithm based on aggregated user performance data
- Develop a community feature allowing peer interaction while preserving the personalised experience
- Create APIs for integration with institutional LMS platforms to fit into existing educational ecosystems

Pebble addresses several critical social issues in higher education:
- Learning Equity: By accommodating diverse learning preferences through multimodal content delivery (visual, auditory, textual), Pebble helps level the playing field for students who struggle with traditional teaching methods. This is especially impactful for neurodivergent learners and those with different cognitive strengths who may be disadvantaged in conventional educational settings.
- Knowledge Accessibility: The platform democratises access to personalised learning experiences that typically require expensive tutoring or specialised programs. By automatically generating tailored explanations, feedback, and study materials from standard course content, Pebble makes high-quality educational support more accessible.
- Engagement Barriers: The adaptive difficulty system and progress tracking directly address motivational challenges that disproportionately affect first-generation college students and those from disadvantaged backgrounds who may lack academic support systems.

Promoting Inclusion and Equity
Pebble enhances inclusion through:
- Multimodal Learning: Supporting different learning styles that may correlate with cultural or neurodevelopmental differences
- Self-paced Progression: Allowing students to master content at their own speed without judgment
- Personalised Feedback: Providing private, constructive assessment that reduces social comparison
- 24/7 Accessibility: Accommodating students with external responsibilities (work, family) who need flexible study times

Impact Measurement Framework
We will measure social impact through:
- Learning Outcome Equity: Comparing performance improvements across different demographic groups and learning preference profiles
- Engagement Distribution: Monitoring usage patterns to ensure equitable benefit across student populations
- Feature Utilisation: Tracking which modalities are most utilized by different user groups
- Survey-based Assessment: Collecting feedback on perceived inclusivity, accessibility, and educational support
- Academic Performance: Measuring improvements in course outcomes, particularly among historically disadvantaged groups

Responsible AI
- Guardrails for all LLM outputs to ensure appropriate content filtering, prevent jailbreaking etc.
- Prompt engineering techniques to maintain educational relevance and accuracy
- Content verification processes to ensure generated explanations and feedback align with course materials
- Human oversight in evaluation metrics to ensure AI systems promote genuine learning outcomes rather than gaming the system
By centering equity and inclusion in both design and measurement, Pebble aims to not just improve learning outcomes but to make quality personalized education more democratic and accessible.

Personal Information Collection Statement (PICS):
1. The personal data collected in this form will be used for activity-organizing, record keeping and reporting only. The collected personal data will be purged within 6 years after the event.
2. Please note that it is obligatory to provide the personal data required.
3. Your personal data collected will be kept by the LTTC and will not be transferred to outside parties.
4. You have the right to request access to and correction of information held by us about you. If you wish to access or correct your personal data, please contact our staff at lttc@eduhk.hk.
5. The University’s Privacy Policy Statement can be access at https://www.eduhk.hk/en/privacy-policy.

• I have read and agree to the competition rules and privacy policy.