KareBud is a deployed health companion built around a simple idea: a useful AI response needs context, not just a prompt box.
The problem
Health information is personal, fragmented, and easy to misread. The product needed to make journaling low-friction while keeping generated guidance grounded in what the user actually shared.
The system
The application uses Next.js, Supabase, FastAPI, and Docker. Voice journals are transcribed with Whisper, then passed through a Critic–Designer–Implementer prompt pipeline:
- The Critic looks for missing context and weak assumptions.
- The Designer shapes a response plan around the available evidence.
- The Implementer produces the final user-facing response.
voice journal -> transcript -> contextual review -> response plan -> final guidance
What I learned
Multi-step prompting is valuable when each step has a clear job. More agents are not automatically better; the pipeline earns its complexity only when it catches assumptions or improves the final response. KareBud currently demonstrates the product and architecture. I do not publish adoption or evaluation claims that have not been verified.