Bachelor Thesis

A full-stack adaptive learning platform developed as a bachelor's thesis, exploring whether deep learning-based knowledge tracing can be meaningfully translated into actionable pedagogical recommendations for classroom teachers. The project serves as a proof of concept for a planned joint study between Berliner Hochschule für Technik (BHT) and a partner institution, investigating the practical applicability of AI-driven learning analytics in real educational settings. At the core of the system is an Attentive Knowledge Tracing (AKT) model — a transformer-based architecture that processes a student's full interaction history across problems and skills to produce continuous mastery probability estimates. Raw AKT output probabilities are post-processed into interpretable skill-level mastery profiles, which are then mapped to a three-tier recommendation framework grounded in established learning theory: Practice (70–90% predicted success) for consolidation, Optimal (50–70%) for the Zone of Proximal Development as described by Vygotsky, and Challenge (30–50%) for stretch tasks — drawing on Bloom's mastery learning principles. The backend is built with FastAPI and SQLAlchemy on SQLite, exposing a fully authenticated REST API with JWT-based auth, class/student management, CSV interaction import with duplicate detection and error reporting, and an AKT inference service loaded as a singleton at startup. The frontend, built with Next.js 14 and shadcn/ui, provides teachers with a dashboard for managing classes and students, importing data, and viewing per-student recommendation reports with configurable visualizations — radar charts, bar charts, or tables — exportable as formatted PDFs via jsPDF and html2canvas. A small user study with three teachers found the mastery transformation concept promising, but raised concerns around AI transparency and explainability — a known challenge in deploying black-box models in educational contexts, and an area identified for future work.

Key Features

• Attentive Knowledge Tracing (AKT) transformer model for continuous per-skill mastery probability estimation.
• Three-tier recommendation framework (Practice / Optimal / Challenge) grounded in Vygotsky's ZPD and Bloom's mastery learning.
• FastAPI backend with JWT authentication, class/student management, and CSV import with duplicate detection.
• Next.js 14 teacher dashboard with configurable report visualizations: radar charts, bar charts, and tables.
• PDF export of per-student recommendation reports via jsPDF and html2canvas.
• User study with three teachers evaluating pedagogical applicability and AI transparency.