MediFlow AI

The Challenge

Building a health app that provides accurate, trustworthy AI-powered health insights while maintaining strict privacy compliance posed multiple layers of complexity. First, the AI needed to be accurate enough to be genuinely useful without crossing the line into medical diagnosis — a fine line that required careful prompt engineering and output filtering. Quebec privacy law (Law 25) and federal PIPEDA regulations imposed strict requirements on how health data is collected, stored, and processed. The app needed to work in both English and French with equal quality, including the AI responses. The appointment booking system needed to integrate with multiple clinic management systems that use different APIs and data formats. Battery efficiency was critical — a health companion app that drains your phone defeats its own purpose. Finally, the AI needed to handle sensitive conversations with appropriate empathy and always err on the side of recommending professional care for anything potentially serious. The client also wanted the app to work partially offline, caching enough health information that users in rural areas with spotty connectivity could still access their medication schedules and basic wellness advice.

Our Solution

We developed a custom RAG (Retrieval-Augmented Generation) system that combines a curated medical literature database with the Claude API for nuanced, empathetic health conversations. The RAG pipeline indexes over 50,000 peer-reviewed medical articles and clinical guidelines, ensuring the AI provides evidence-based responses rather than hallucinating medical advice. Every AI response passes through a medical safety filter that flags and softens any language that could be interpreted as a diagnosis, always appending appropriate disclaimers and professional consultation recommendations. The app was built with React Native for cross-platform deployment, with native modules for HealthKit (iOS) and Google Health Connect (Android) integration. End-to-end encryption protects all health data in transit, while PIPEDA-compliant data storage on AWS Canada (Montreal region) ensures data sovereignty. The bilingual system uses language detection with locale-aware medical terminology, so the AI naturally responds in the user language with correct medical terms in that language. For offline functionality, we implemented a smart caching system that stores the user medication schedule, recent health insights, and essential first-aid information locally using encrypted SQLite. The appointment booking module uses a middleware layer that normalizes different clinic API formats into a unified interface. The development process included consultation with two practicing physicians who reviewed AI outputs for medical accuracy throughout the 8-month build cycle.

Results

Tech Stack