As attention-related cognitive disorders affect millions globally, there is a critical need for accessible, human-centric tools to monitor and improve mental well-being. This paper introduces NOVA, a real-time neurofeedback system that integrates electroencephalography (EEG) with local Large Language Models (LLMs) to enhance cognitive focus and relaxation. Addressing the ergonomic challenge of human-computer interaction in digital health, NOVA utilizes the Muse-2 EEG headset to capture brainwave data, which is processed through bandpass filtering (1–30 Hz) and Power Spectral Density (PSD) feature extraction. The system employs a fine-tuned, 2-billion parameter Gemma-2 model running locally to ensure data privacy and a response latency of less than 0.2 seconds, significantly outperforming cloud-based alternatives like GPT-4o. A custom adaptive interface scales difficulty based on the user's real-time focus levels, providing an ergonomic feedback loop. Clinical testing over ten sessions demonstrated substantial increases in focus and relaxation durations, with findings validated by strong correlation (ρ > 0.9) and high statistical significance (p < 0.001). The fine-tuned local model achieved 97% accuracy in state classification, proving the efficacy of using consumer-grade hardware for high-fidelity cognitive intervention. This research contributes to a scalable, privacy-preserving framework for healthcare informatics and personalized ergonomics, demonstrating how integrated neuro-AI systems can fundamentally improve human performance and system resilience in healthcare environments.