Privacy-Oriented On-Device Medical Document Understanding for Elderly Users Using a Lightweight Language Model

I developed Vault Guard as an iOS-based AI application to help people, especially elderly users, understand information in medical documents more easily. The main idea behind the project came from a simple problem: medical prescriptions and documents can be difficult to read, and the information in them is not always easy for everyone to understand. In the application, I use Apple's Vision framework to extract text from a prescription or other medical document. The user can either take a picture using the device camera or select an existing image. After the text is extracted, I pass it to a small Qwen2.5-0.5B-Instruct 4-bit language model that runs locally on the device using the MLX framework. The model then converts the extracted information into a shorter and easier-to-understand explanation. I also added voice interaction so that users can communicate with the application using speech and listen to the generated response through text-to-speech. The application includes multiple language options because I wanted the system to be useful for people who may not be comfortable using English. Another important part of my design is local AI processing. Instead of depending completely on a cloud-based AI service, the language-model processing is performed on the Apple device. I designed Vault Guard as a document-understanding and accessibility tool, not as a medical diagnostic system. My goal is to make existing medical information easier to access and understand, while exploring how small AI models can be used directly on mobile devices. As the next stage of the project, I plan to evaluate the system using OCR accuracy, explanation quality, response time, memory usage, multilingual performance, and privacy-related measurements.