Afya-Yangu AI – Offline Clinical Decision Support for Level 2 & 3 Care

1.Problem & Context

Why Afya-Yangu AI?

Kenya’s level 2 and 3 facilities (dispensaries and health centres) are the backbone of primary care, yet clinicians often have limited time and tools to search and interpret national guidelines during a busy shift.

Guidelines exist for child health, maternal and newborn care, TB, HIV, malaria, and NCDs – but they are scattered across PDFs, booklets, and static websites.

The result is:

• Variation in care between facilities and providers.
• Overuse or misuse of antibiotics.
• Missed danger signs and delayed referrals.

Afya-Yangu AI aims to bridge this gap by making Kenyan guidelines “conversational” and available offline at the point of care.

2.Technical Architecture (High-Level)

Under the Hood

1. Model: MedGemma SLM, fine-tuned on:

• Synthetic and real QA pairs from Kenyan guidelines.
• Common level 2 & 3 clinical scenarios.

2. Knowledge Base:

• Kenyan level 2 & 3 guidelines converted from PDF to text.
• Chunked and tagged by topic (e.g. child, maternal, HIV, TB, NCD).
• Embedded and stored in a FAISS index for efficient similarity search

3. RAG Pipeline:

• User query → FAISS retrieves top-k relevant chunks → combined into prompt → MedGemma generates answer.
• Deployment:
• Containerised (e.g. Docker) for on-prem or local server.
• Simple REST API for mobile / web front-ends.

4. Governance & Safety:

• Strict prompting: no advice outside Kenyan guidelines; defer or refer when out of scope.
• Logging of queries/answers (de-identified) for quality improvement and audit.

3. Impact & Roadmap

Expected Impact

• Improved guideline adherence at level 2 & 3.
• More appropriate antibiotic use.
• Better triage and referral decisions, especially for maternal and child health.
• Data to understand what clinicians struggle with (through anonymized query logs).

Roadmap bullets:

1. Phase 1 – Prototype with a subset of guidelines (e.g. child health & maternal care).
2. Phase 2 – Expand guidelines, optimise latency on low-power hardware.
3. Phase 3 – Pilot in select facilities, collect usability & clinical outcome signals.
4. Phase 4 – County-level scaling and integration with EMRs / DHIS2 data flows.