Many foundational studies have highlighted how AI can improve education, from automated grading to adaptive, real-time instruction. However, these benefits remain inaccessible to millions of students due to digital inequality. Traditional AI grading systems depend on constant cloud connectivity and high-power hardware, creating barriers for schools with limited electricity or minimal technological infrastructure. We present a proof of concept for an offline, browser-based educational assessment system that leverages on-device AI to provide automated grading without requiring internet connectivity or cloud infrastructure. The system combines WebAssembly-compiled optical character recognition with Google’s Gemini Nano language model, embedded directly in the Chrome browser, to process handwritten student submissions locally on a user’s device. This work grounds the discussion of offline, privacy-preserving educational AI in a proof-of-concept implementation. We also reflect on the ethical and societal implications of automated grading, including the risk of automation complacency and the limits of offline access in addressing educational inequity.