This paper presents an AI-assisted virtual keyboard controlled entirely through eye gaze and blink detection, offering a fully hands-free typing interface for users with motor impairments. The system integrates MediaPipe for eye landmark tracking, OpenCV for real-time gaze estimation, and a Tkinter-based graphical interface for virtual keyboard interaction. Users can type via sustained gaze (dwell selection) or blink-triggered actions, with an adaptive debounce mechanism to prevent accidental inputs. To enhance typing fluency, a dual-layer language model is implemented: a lightweight bigram model derived from the Brown corpus for offline prediction, and an optional transformer-based GPT-2 engine for contextual and semantically rich suggestions. Experimental evaluation with two participants demonstrated high blink classification accuracy (94.6%), stable gaze-to-cursor mapping (91.3%), and improved typing speed (up to 9.4 WPM) during AI-assisted sessions. Visual plots of cursor smoothing, blink timelines, and accuracy breakdowns validated system responsiveness and usability. The interface was rated intuitive and fatigue-free, with participants preferring blink overrides and GPT suggestions for longer phrases. The proposed model offers a cost-effective, real-time communication tool that bridges computer vision and AI for assistive technology, operating entirely on standard webcams without proprietary hardware.