Case Study · 2026
MamaCare Triage
An offline-capable, multimodal AI triage assistant for rural clinics, built for the Build with Gemma: GenAI for SDGs hackathon. It returns a condition/urgency assessment, a safe medication dosage, and danger-sign alerts — translated into Yoruba for community health workers.
MamaCare Triage — 2026
The Problem
Rural clinics triage pregnant women with no doctor in the room
In many rural facilities, the first — and sometimes only — health worker a pregnant woman meets is a community health worker (CHW) with limited diagnostic support. Referral decisions are made on instinct. Danger signs get missed, and the nearest doctor is hours away.
- ✗No specialist available when a patient presents with symptoms
- ✗Medication dosing errors are life-threatening with no physician to verify
- ✗SDG 3 target: reduce maternal mortality — triage is the first link in that chain
- ✗Connectivity can't be assumed at the point of care
Why this solution
Gemma's function-calling, forced into a safe clinical structure
Instead of free-form chat, MamaCare forces Gemma through native function-calling into a strict clinical schema: condition, urgency tier, dosage suggestion, danger-sign flags. Structured output is what makes an LLM usable in a clinical workflow — the model can't ramble when the schema won't let it.
- ✓Multimodal: Gemma 4 reasons over symptoms AND images where available
- ✓Function-calling = guardrails: outputs must fit the clinical schema
- ✓Offline-capable design for clinics with intermittent power/connectivity
- ✓Yoruba translations meet the CHW in their working language
How it works
01
Intake
CHW enters symptoms (and optional photo) in the clinic interface
02
Reasoning
Gemma 4 multimodal model reasons over the presentation
03
Structured triage
Function-calling forces condition + urgency tier + danger-sign flags
04
Dosage check
Safe medication dosage surfaced with alerts for contraindicated signs
05
Localisation
Output translated to Yoruba for the CHW to act on immediately
Challenges I faced
⚠️ LLMs hallucinate — unacceptable in health
→ fixed: Schema-locked function-calling plus explicit danger-sign alerts; the tool errs on the side of referral
⚠️ Hackathon timebox
→ fixed: Scope locked to the triage loop: intake → structured assessment → translated output. No auth, no billing — clinical core only
⚠️ Connectivity at point of care
→ fixed: Offline-capable architecture designed from the first commit
⚠️ Yoruba medical phrasing
→ fixed: Curated translation layer for clinical terms rather than raw machine translation
Results
SDG 3
aligned — maternal health
3
structured outputs per triage
Yoruba
output language for CHWs
100%
schema-locked responses
🧰 Built with
💡 What it taught me
In health AI, the guardrails ARE the product. Function-calling schemas, danger-sign alerts and fail-to-referral defaults matter more than model quality.