Zeya | Devpost

Inspiration

Maternal mortality remains a critical challenge in low-resource settings. In Migori County, Kenya, pregnant women often lack access to timely, culturally appropriate antenatal education. Many cannot attend regular clinic visits due to distance, cost, or time constraints, yet they carry mobile phones with WhatsApp installed.

We were inspired to leverage Google's Gemini API to create an AI-powered WhatsApp chatbot that brings expert antenatal guidance directly to expectant mothers' phones—available 24/7, in both English and Swahili.

What it does

Zeya is an AI-powered maternal health assistant accessible via WhatsApp. It provides:

Personalized antenatal education based on gestational age
Danger sign detection with automatic emergency alerts
Database-driven facility referrals for local hospitals in Migori County
Bilingual support (English & Swahili)
3-step registration (consent → name → gestational age)
Contextual AI responses using Gemini 3 Flash Preview with WHO/Kenya MOH guidelines

When a user sends a message, the system:

Detects 8 categories of obstetric danger signs using regex patterns
If detected, immediately provides emergency contacts for nearby verified health facilities
Otherwise, generates contextualized educational content via Gemini API with conversation history

How we built it

Backend (Python/FastAPI)

Gemini 3 Flash Preview (gemini-3-flash-preview) for AI response generation
WhatsApp Cloud API for messaging
PostgreSQL with async SQLAlchemy for data persistence
Redis for conversation history caching (24-hour TTL)
FastAPI with gunicorn + uvicorn workers for production deployment
Alembic for database migrations

Database Schema

Users, conversations, health facilities, engagement metrics, knowledge assessments
PostgreSQL enums for study groups, facility types, admin roles

AI Architecture

System prompt engineered with WHO antenatal care guidelines and Kenya MOH protocols
Dynamic gestational age context injection
Conversation history (last 6 turns) from Redis
Safety constraints for medical advice boundaries

Admin Dashboard

JWT authentication with role-based access control
Health facility management (CRUD operations)
User analytics and data export (CSV/Excel)
Emergency contact configuration per county

Infrastructure

Docker Compose multi-container setup
ngrok for webhook tunneling during development
Non-root container users for security
Deep health checks (DB + Redis)

Challenges we faced

Enum Serialization Bug: Pydantic 2.x wasn't serializing enum values correctly for PostgreSQL. Spent hours debugging why StudyGroup.INTERVENTION was being passed as "INTERVENTION" instead of "intervention". Solution: Explicitly convert enums to their .value in service layer.
Emergency Response Timing: Initially, danger sign detection happened after AI generation, causing delays. Refactored to detect-first architecture for faster emergency alerts.
Database-Driven Facilities: Migrated from hardcoded hospital contacts to a flexible health facility database with county-based lookups, priority sorting, and verification flags.
Context Window Management: Balancing conversation history length (6 turns) with token limits while maintaining coherent responses.
Production Hardening: Securing the application for deployment—switching from development mode to gunicorn, adding non-root users, protecting admin endpoints, implementing pre-commit hooks for secret scanning.

What we learned

Gemini 3 API capabilities: The gemini-3-flash-preview model excels at generating medically accurate, culturally sensitive responses when given structured prompts
Prompt engineering for health: Careful system prompt design is critical for maintaining safety boundaries in medical advice
Async Python at scale: SQLAlchemy async patterns, Redis caching strategies, and FastAPI dependency injection
WhatsApp webhook patterns: Handling message deduplication, status callbacks, and retry logic
Enum handling in SQLAlchemy + Pydantic: The nuances of PostgreSQL enum types with Python type systems