Designing for a sensitive topic
Menstrual health questions are shaped by stigma, privacy, language, and access—not just information retrieval. We used a familiar WhatsApp interface to make expert-grounded guidance approachable in a low-resource setting.
- Designed around an existing communication habit rather than a new app.
- Focused on culturally sensitive, low-friction interactions.
- Used the study context to shape the system rather than treating deployment as an afterthought.
Grounding over improvisation
The assistant retrieves from an expert-curated knowledge base before composing an answer. The goal was to reduce unsupported responses and make the system useful for questions where casual hallucination would be harmful.
- Built the retrieval-augmented response pipeline in Python.
- Deployed the backend on AWS EC2 with DynamoDB for application state.
- Kept the WhatsApp channel separate from the knowledge and response layers.
Field evaluation
The chatbot was evaluated during a two-week deployment that produced 403 messages. The qualitative and interaction evidence showed how an expert-grounded LLM could become a trusted health resource in a low-resource setting.
- Evaluated behavior in real use rather than only offline prompts.
- Studied trust and stigma alongside technical performance.
- Published the work at ACM CHI 2026.