Healthcare is one of the highest-stakes environments to introduce AI. People's decisions about their health depend on the guidance they receive — and that guidance needs to be clear, safe, and grounded in real information. Here is how we approached building Carebrij's AI-assisted symptom assessment.
The core challenge with Carebrij was not the model selection. It was the trust problem. Medical guidance from an AI raises an immediate question: can this be wrong? Our answer was to build a system that guides, not diagnoses. The AI helps users understand patterns in their symptoms and connects them to verified professionals — rather than rendering a medical verdict.
I contributed to the LLM integration layer using FastAPI and Python. The system takes symptom descriptions, applies structured prompt engineering to extract medically relevant signals, and surfaces a ranked set of professional categories that match the user's input against a verified provider database.
One of the most important decisions was how to handle uncertainty. Rather than returning a confident but potentially wrong answer, the system expresses calibrated uncertainty — 'this pattern is consistent with...' rather than 'you have...' — and always recommends professional confirmation.
We used RAG-style retrieval to ground responses in real medical category data rather than relying purely on the model's parametric knowledge. This kept outputs relevant to Nigeria's healthcare context, which has specific provider types and local realities that global models don't naturally account for.
Impact beyond the screen
- Delivered an AI-assisted symptom flow that guides rather than diagnoses.
- Integrated FastAPI services with a React frontend for real-time assessment.
- Applied RAG and structured prompt engineering for medically grounded outputs.
- Designed uncertainty-aware responses that always recommend professional confirmation.
Lessons from the work
- In high-stakes AI, expressing calibrated uncertainty builds more trust than false confidence.
- Domain context — local healthcare realities — must shape model behaviour, not just training data.
- The user interface matters as much as the model: how results are displayed shapes how they are understood.
- Grounding responses in verified, real-world data dramatically reduces hallucination risk.
