Page 95 - AI for Good Innovate for Impact

AI for Good Innovate for Impact

(continued)

Item Details
Model Training and – LLM Integration: GPT-4 via OpenAI for response generation
Fine-Tuning – Retrieval System: Pinecone vector database for RAG (Retriev- 4.1-Healthcare
al-Augmented Generation)
– Hosting & Infrastructure: Managed via DigitalOcean, using
Docker for containerization
– Speech Capabilities:

• Text-to-Speech (TTS) via ElevenLabs—enabling audio responses
• Speech-to-Text (STT) via Whisper V3/Replicate—enabling audio
input from users

– Monitoring & Evaluation:
• METABase: custom dashboard for chatbot performance and
engagement analytics
• Flower: backend tool for real-time technical performance moni-
toring
– Content & Persona Delivery: Synthesia avatars are used to
deliver health content in video form

Testbeds or Pilot Deploy- Active, with ongoing deployment and iterative refinement. (Leba-
ments non and Jordan)

Code repositories Not Available

2 Use Case Description

2�1 Description

Health misinformation, language barriers, and inequitable access to trustworthy health
information continue to threaten community well-being in low- and middle-income countries
(LMICs), especially among displaced and vulnerable populations [1]. These barriers are often
exacerbated by low literacy levels, social stigma around health topics, and a lack of culturally
sensitive communication platforms. To address these issues, HIBA (Health Information Bot
Assistant) was developed as a WhatsApp-based, AI-powered chatbot that delivers interactive,
multilingual health messaging in both text and audio formats [9][10]. Its key objective is to
improve equitable access to validated health information by leveraging AI tools such as LLMs,
STT, and TTS to enhance inclusivity and user engagement. HIBA’s solution architecture is based
on a RAG framework integrated with ReAct prompting, enabling the chatbot to deliver accurate,
relevant, and user-friendly responses while preventing hallucinations [2][3].
HIBA has been piloted in Lebanon and Jordan with refugee and host communities, integrating
feedback from multi-phase user testing to optimize content flow, accessibility, and tone. The
chatbot supports 28 languages, including Arabic dialects, and applies community-validated
ethical guardrails to ensure that content is informational and not diagnostic. The system
architecture incorporates Pinecone for vector search, ElevenLabs for human-like TTS, and GPT-4
for language understanding. For STT, the open-source model Whisper V3 from OpenAI was
utilized, which worked well with multiple Arabic dialects, in conjunction with an added layer of
LLM inference that tries to match the utterances to intents. As part of the HIBA development
process, a comprehensive four-level testing framework was implemented to ensure the quality,

59