Empowered by Jacaranda Health’s PROMPTS platform, a mother-to-be confidently accesses the guidance and support she needs—right at her fingertips.
Jacaranda Health (JH) is pioneering the use of generative AI to transform how underserved mothers across Sub-Saharan Africa access, understand, and act on vital maternal and newborn health information. Since 2016, Jacaranda has partnered with national and county governments, frontline nurses, and mothers to address persistent gaps in public hospital care, where most underserved women receive services. Through a combination of digital tools and a robust data infrastructure, Jacaranda enables health system managers and providers to identify and address quality gaps in maternal and newborn care—improving health outcomes and experiences at every layer of a mother’s support system.
The cornerstone of this effort is PROMPTS, a digital health platform that has reached over 2.77 million mothers and newborns to date. Originally designed around low-cost SMS, PROMPTS nudges women toward timely health-seeking behaviors, delivers credible stage-specific information, and leverages a customized Large Language Model (LLM), known as UlizaMama, to power an AI-enabled clinical helpdesk. Through this helpdesk, mothers receive rapid, personalized responses to their questions in English and Swahili. If a risk is identified, trained nurses are seamlessly engaged to guide and expedite referral to appropriate care. PROMPTS also collects anonymous, real-time feedback from mothers on their care experiences, empowering health system managers to continuously improve quality and respectfulness of care.
Building on the proven PROMPTS system, Jacaranda now seeks to leverage voice-based AI solutions to reach mothers with limited literacy, sight impairments, or other barriers to text-based engagement. Drawing on OpenAI’s technology, Jacaranda will adapt its LLM to provide a voice Q&A service that can overcome language and literacy constraints. This voice functionality will allow mothers to call a toll-free number, ask questions in Swahili (or code-mixed English-Swahili), and receive timely, culturally-appropriate spoken responses. Whenever a danger sign emerges, a referral can be activated by a nurse, just as with SMS, ensuring equitable and inclusive care guidance.
Since Swahili is the preferred language for 60% of PROMPTS users, it is essential that the voice solution functions reliably and efficiently in Swahili. Jacaranda Health plans to refine and deploy a Swahili-language voice Q&A system that can rapidly and accurately provide information, encourage health-seeking behaviors, and facilitate referrals.
Jacaranda will pursue three main objectives to ensure the systems success.
Initial tests revealed that code-mixed queries and dialectal variations complicate STT and TTS accuracy for Swahili. To address these challenges, Jacaranda Health will fine-tune OpenAI’s Whisper models using diverse datasets that capture a range of Swahili accents, dialects, and conversational styles typical of PROMPTS users. Training data sources will include:
Additionally, the team will create a 20-hour dataset of high-quality, text-based responses generated by the LLM. Thirteen helpdesk agents will record these sentences to improve TTS output, ensuring natural, contextually appropriate Swahili that aligns with users’ linguistic preferences.
After developing the improved STT and TTS models, Jacaranda Health will integrate them into PROMPTS. Initial tests were conducted using Telegram’s free, scalable API, but the finalized solution will operate through standard phone lines, removing the need for a smartphone or data connection. The upgraded voice functionality will be connected to UlizaMama, the customized LLM driving PROMPTS.
Jacaranda Health will conduct qualitative user testing with a diverse group of mothers to determine product requirements, including ideal call timing, message length, and frequency. Previous IVR testing showed that shorter messages and certain times of the day (weekend mornings, weekday evenings) increased engagement. Insights from these tests will inform the design and rollout of the Swahili voice Q&A solution.