SocratiQ: A Generative AI-Powered Learning Companion for Personalized Education and Broader Accessibility

AI-Powered Learning Enhancement

Context: Online machine learning systems textbook (CS249r course at Harvard University) and broader STEM fields, and advanced college-level curricula.

Implementation: Utilizes language models (LLMs) to enhance the learning experience through various methods. This includes: (1) Integration into the reading experience to generate quiz questions and engage in natural language conversations (SocratiQ), implemented as a client-side application that injects itself into the webpage, extracting text, indexing it, and injecting quiz buttons. Uses serverless processing and API calls to language models. (2) Providing explanations tailored to the student's level using a slider menu for different difficulty levels (Personalized Explanations). (3) Dynamically generating assessments by indexing textbook content, adding buttons after sections, and passing the associated text to the LLM to create quizzes with multiple-choice options and answer explanations (Adaptive Assessments). (4) Emphasizing textbook material as the primary focus of the language model through in-context prompts, using algorithms to identify similar paragraphs to the user query (Bounded Learning).

Outcomes: Enhanced student engagement, personalized learning pathways, and broader accessibility to high-quality education. Improved comprehension, retention, and critical thinking skills. Increased student interaction with textbook material and assessment completion. Adapt explanations to align with a student’s background knowledge and address specific areas of difficulty. Ensures that learners receive accurate, focused, and contextually relevant information. Provides multiple opportunities to test knowledge and focus on areas of weakness.

Challenges: Context retention, personalization, cost of API calls. Effectively adjusting the complexity and depth of explanations to align with student's background knowledge. Generating questions relevant to the selected content, at the appropriate academic level, and in the correct format. Balancing the curated textbook content with the supplementary knowledge from the language model’s training data. Maintaining the accuracy, depth, and coherence of the primary textbook while still having access to broader supplementary information when needed.

Gamification

Context: Online machine learning systems textbook (CS249r course).

Implementation: Implemented five features designed to motivate learners: Progress, Streaks, Passing quiz attempts, Badges, Engagement Heatmap.

Outcomes: Enhanced user engagement and provided opportunities for self-assessment. Motivated learners and provided a sense of accomplishment.

Challenges: Mixed responses to gamification features. Some students found them motivating while others had a more nuanced view.

Implementation Barrier

Integrating AI tools effectively into existing educational frameworks and mitigating the potential erosion of human connections.

Proposed Solutions: Careful consideration of how to preserve and enhance traditional teaching approaches. Leveraging AI’s capabilities for personalization and efficiency while maintaining the human connections and community support essential for meaningful learning experiences.

Technical Barrier

Token limitations and real-time requirements, which constrain the amount of text that can be processed in a single request, particularly during quiz generation. Difficulty in teaching complex technical content. Machine learning systems courses require students to simultaneously grasp concepts that span data engineering, model architecture, and computational infrastructure.

Proposed Solutions: Employing a selective text inclusion strategy (vectorizing only the first k sentences) and question caching with a reuse strategy to optimize system performance. Future AI learning companions will need sophisticated mechanisms to dynamically adjust their depth of explanation.

Cost Barrier

API calls to language models introduce an associated cost. Open access policy potentially generates a large volume of requests.

Proposed Solutions: Multi-model, multi-service strategy. Use Mixtral-8x7b and Gemini as primary AI services, and use other services as a backup. Use question caching and reuse strategy to reduce API calls.

Privacy Barrier

Extensive data collection and implementation barriers such as limited understanding of data-driven systems and questions of data sovereignty.

Proposed Solutions: Local-first architecture, storing all chats securely in the browser’s IndexedDB, and quiz statistics are maintained in local storage, minimizing external data dependencies. Selective centralized caching system. Secure progress sharing through cryptographically hashed PDFs.

Project Contact: Dr. Jianhua Yang

LLM Model Version: gemini-2.0-flash-lite