GIST, South Korea · Spring 2025
AI Wellness Assistant
Instead of one model behind one long prompt, three small agents that each do one job, and know how to find each other.
The idea
People describe health worries messily: a headache, bad sleep and a skipped lunch, all in one sentence. This assistant splits the question up and sends each part to a specialist agent.
Built during my exchange semester at GIST, on Google’s Agent2Agent (A2A) protocol while its tooling was still very new.
How it works
- Three independent agents, each running as its own JSON-RPC server: an orchestrator, a symptom checker, and a lifestyle recommender for sleep, diet and daily habits.
- The orchestrator doesn’t hard-code its helpers. At startup it reads a registry and fetches each agent’s card from
/.well-known/agent.json. Adding a fourth agent means adding a URL, not editing the router. - Two ways in: a Flask web chat and a command-line client.
Python throughout, with Google’s Agent Development Kit and Gemini for the agents, FastAPI and Uvicorn per service, and Pydantic models for every message.
Result
The routing held up on mixed input, correctly splitting a single complaint across both specialists. The project shipped with two recorded demos and a business justification report.
It gives general wellness information only, not medical advice.