MediGuide: One-Click Medical AI

A Chrome extension powered by Gemini 2.5 Flash and MongoDB MCP that instantly analyzes patient symptoms and fetches medical guidelines.

Comment

Inspiration

Looking up medical guidelines during consultations or when checking personal health symptoms takes too much time. I wanted to build a fast, one-click tool right in the browser. The goal was to create an AI assistant that can instantly analyze symptoms and pull up highly relevant clinical data without disrupting the user's workflow.

What it does

MediGuide is a Chrome Extension that acts as an on-demand medical AI assistant. Users can type their symptoms (e.g., "heartburn and acid reflux") or click quick-tags. The agent then dynamically searches a MongoDB database to retrieve standard clinical guidelines and outputs a clean, professional Markdown report. It also features a seamless real-time toggle between English and Korean UI/outputs.

How we built it

The core brain is powered by Google Cloud Agent Builder and the latest gemini-2.5-flash model. To connect the Chrome Extension to Google Cloud, I built a lightweight backend using FastAPI. The most critical technical achievement was building a custom MongoDB MCP (Model Context Protocol) server. This allows the Gemini model to use the database as a direct tool, autonomously searching and retrieving medical guidelines based on the patient's symptoms.

Challenges we ran into

I hit a major roadblock early on with a "404 NOT_FOUND" API error because the older Gemini models weren't supported in my specific Google Cloud environment. I had to quickly debug the model compatibility and migrate the entire architecture to the gemini-2.5-flash model.

Another big challenge was the search accuracy. Initially, if a user typed a long, natural sentence, the database search failed to find an exact match, causing the AI to return a generic fallback answer. I had to rewrite the MCP search logic on the fly to tokenize the natural language input and perform a multi-keyword OR search, which pushed the matching accuracy to near 100%.

Accomplishments that we're proud of

Getting the MCP server to perfectly bridge Gemini with MongoDB was a huge win. I'm also really proud of the resilience of the system—it handles API errors gracefully and delivers results in just a few seconds. The custom i18n implementation for the English/Korean toggle in the Chrome extension also works flawlessly without needing page reloads.

What we learned

I learned how to properly implement the Model Context Protocol to give LLMs structured, real-time access to a NoSQL database. I also gained a lot of hands-on experience orchestrating Google Cloud AI platform APIs with a Python backend.

What's next for MediGuide

The next step is to expand the MongoDB database with real-world, peer-reviewed medical datasets. I also plan to integrate multimodal capabilities so users can upload images (like skin rashes) directly through the extension for the AI to analyze alongside the text symptoms.

Built With

Share this project:

Updates