Nia

Problem & context

General-purpose chatbots are a poor fit for health questions: they can hallucinate, overstate certainty, and miss cultural context. For a women’s health product, a wrong answer isn’t a bug—it’s a harm. Nia needed to be helpful, warm, and verifiably grounded.

My role

I designed and built the system end-to-end: the retrieval pipeline, the FastAPI service, the conversation design, and the integration into the Check Me platform.

Technical approach

Nia is a retrieval-augmented generation system. Instead of letting a language model answer from its own weights, every question first retrieves relevant passages from a curated corpus of trusted medical content. The model then answers from that context, and responses carry their sources.

• Curated medical-document corpus as the single source of truth
• LangChain retrieval pipeline with OpenAI models for generation
• FastAPI async service so conversations stay responsive on slow connections
• Safety-conscious response design: educational framing, no diagnosis, clear escalation to professionals
• Conversational tone tuned to be warm and culturally sensitive rather than clinical

Decisions & tradeoffs

Grounding constrains the model—Nia declines to speculate beyond its sources, which occasionally means saying “I don’t know, and here’s who to ask.” We chose that tradeoff deliberately: in health, a bounded honest answer beats an unbounded confident one. The conversation design also routes anything that resembles a symptom assessment toward real clinicians via Check Me’s referral workflows.

Outcome & impact

Nia ships inside Check Me, serving its community of 10,000+ users as the platform’s always-available first point of contact for health questions—grounded, cited, and designed to know its limits.

Lessons learned

• Retrieval quality matters more than model choice—garbage context produces confident garbage.
• Designing what an AI won’t say is as much engineering as what it will.
• Warmth and safety aren’t in tension; the most trusted answer is both kind and cited.

Links

• Get the app