MedWise

Inspiration

Claude Health and ChatGPT Health offer exciting ways to get personalized health information, analysis, and insights. However, many consumers do not feel comfortable sharing their personal health information (PHI) with large companies who may not fully back up their promises of data privacy and security with the right policies and actions.

What it does

MedWise is a privacy-first AI health assistant that analyzes your medications, conditions, lab results, and genetic data locally on your device. When external knowledge is needed, it decomposes queries into abstract sub-questions, enforces strict exposure budgets, shows you exactly what data would be sent, and requires explicit approval before sharing anything. Your genetic data never leaves your device—period.

How we built it

We built a local-first system using Python and Streamlit that ingests medical data via CSVs, converts it to structured clinical notes, and performs medical reasoning using local LLM inference via MedGemma. For optional remote calls, we implemented query decomposition with data minimization algorithms, exposure scoring, and real-time payload previews. We also built a pharmacogenomics preprocessing pipeline that converts raw genetic phenotypes into actionable clinical summaries without cloud processing.

Challenges we ran into

Running a 54GB medical AI model locally required quantization and GPU optimization, while balancing privacy with utility meant building a compartmentalization system that shows users exactly what gets shared. Making privacy tangible through exposure scores and payload previews was harder than the technical implementation itself.

Accomplishments that we're proud of

We proved sophisticated medical reasoning can run entirely on a laptop without cloud dependency, and built a privacy system that’s actually usable with visual exposure tracking and payload previews. We’re one of the few projects integrating pharmacogenomics properly and making data privacy a top priority.

What we learned

MedGemma 27B knows incredible amounts of pharmacology but needs prompt engineering to balance caution with actionable insights, and privacy features need just as much UI/UX love as functionality. Drug interactions are far more complex than we expected, reinforcing that this is decision support, not diagnosis.

What's next for MedWise

Short-term: mobile apps with on-device ML and EHR integration; medium-term: clinical validation study and curated interaction database; long-term: FDA regulatory pathway and provider partnerships. We’re proving medical AI doesn’t have to sacrifice privacy to save lives.​​​​​​​​​​​​​​​​

Built With

• cursor
• hipaa
• medgemma
• minimax
• nist
• opus