The Cognitive Atelier: Emotion-Aware AI Tutoring Through Large Language Models and Adaptive Learning

This paper presents Lumina, a full-stack emotion-aware intelligent tutoring system (ITS) that couples webcam-based facial emotion recognition (FER) with a multi-provider large language model (LLM) orchestration layer to adapt tutoring tone, content depth, and feedback style to a learner’s inferred affective state. The system is implemented as a Next.js frontend and a Flask/SQLAlchemy backend, and integrates an off-the-shelf convolutional FER model with a rule-based emotion-to-tonemapping layer, a cascading LLM “neural link” that falls back across Gemini, Groq-hosted Llama-3.3-70B, and a deterministickeyword-matching retriever, and a research-aggregation module that combines Wikipedia summaries, video references, and LLM-generated lessons. Rather than reporting fabricated benchmark numbers, this paper documents the system’s architecture in detail and reports a small, fully disclosed component-level evaluation executed directly against the deployed source code: webcamframe-to-emotion inference latency under CPU-only inference, correctness of the “no-face” fallback path, and an ablation of the system’s JSON-repair heuristic used to stabilize LLM-generated structured output. Engineering trade-offs are discussed throughout, including the use of a static heuristic mapping instead of a learned policy, single-frame (non-temporal) affect inference, and concrete security limitations identified in the codebase (hardcoded fallback API credentials and secret keys).Lumina is positioned relative to recent work on affective and LLM-driven tutoring systems, and a research agenda toward dataset-grounded evaluation of its pedagogical effectiveness is outlined.