2026 ASEE Annual Conference & Exposition

Traditional computer science education platforms often provide generic feedback that fails to address individual student learning gaps, particularly in complex subjects like database systems and SQL programming. This paper presents the design, implementation, and evaluation of an intelligent educational platform that combines Large Language Models (LLMs), knowledge graphs, and adaptive feedback systems to provide personalized learning experiences for database systems courses.
Our system addresses three critical challenges in CS education: (1) providing contextually appropriate feedback without revealing solutions, (2) tracking student progress across interconnected concepts, and (3) enabling instructors to identify class-wide learning patterns. The platform integrates a PrairieLearn-compatible assessment interface with a fine-tuned GPT-4o-mini model for semantic error classification, automatically categorizing student mistakes into seven distinct SQL error types (JOIN, GROUP BY, WHERE, alias errors, Misuse of Aggregate Functions, Subquery Logic Errors, and Ambiguous Column References).
The core innovation lies in our knowledge graph architecture built on Neo4j, which automatically extracts conceptual relationships from course materials using LLM-powered document processing. This graph enables fine-grained progress tracking by mapping student submissions to specific SQL concepts and their prerequisites. Students receive personalized insights after five attempts, highlighting their most frequent error patterns and suggesting targeted practice areas. The system generates adaptive practice problems using templated question generation informed by the knowledge graph structure.
For instructors, we developed a comprehensive dashboard that visualizes class-wide performance patterns, concept mastery levels, and common error clusters. The system processes student submissions through vector embeddings stored in PostgreSQL, enabling similarity-based analysis of common mistakes and near-correct solutions. Our evaluation demonstrates the system's effectiveness in identifying learning gaps and providing actionable insights for both students and instructors.
The platform has been deployed and tested in CS 411 (Database Systems) at the University of Illinois, processing over 1,000 student submissions across multiple assessment types. Results show improved student engagement through personalized feedback and enhanced instructor awareness of class learning patterns. This work contributes to the growing field of AI-assisted education by demonstrating how knowledge graphs can bridge the gap between individual learning needs and scalable educational technology.