A health NGO considered a chatbot for community health workers asking common medical questions. They reimagined by adapting an existing open-source tool rather than building new.
Forked the WHO/Praekelt openCHW knowledge assistant and translated content for local protocols.
Community health workers in rural Tigray and the Amhara region. Most are women who walk 6-15 km between households. They carry paper protocols, often outdated. When a mother asks about a feverish infant, they sometimes guess. The nearest clinical supervisor is a half-day's travel away.
Christ would see that these workers are bearing weight that the system never trained them for. He'd see that the gap between their responsibility and their training is being patched by their own conscience — and that this is unsustainable. He'd see that the answer is not more burden on them, but better support around them.
You're naming a real gap and a real risk: well-meaning workers making clinical guesses in isolation. A chatbot will not replace a supervisor, but it might give workers a reference they currently don't have. The danger is mission creep — a 'reference tool' that becomes a 'diagnostic tool' because the worker has no one else to ask. Whatever you build, build it small, and keep human escalation in the loop.
A smartphone-accessible chatbot that answers common clinical questions for community health workers, trained on WHO IMCI protocols and local Ministry of Health guidelines, available in Tigrinya, Amharic, and English.
Almaz, a CHW serving 12 villages in Tigray, currently carries a 280-page paper protocol manual she can't search. When she sees a sick child, she has 4 minutes to remember the right action. We have ~1,800 CHWs in the same situation. The current alternative is no reference at all.
Partially. A supervisor in every cluster would be better. We don't have the budget. A printed quick-reference card would help but it's already been tried and discarded — they get lost or wet.
Real need: just-in-time, searchable reference. Not diagnostic AI — that would exceed both their training and our legal scope. We want a smarter manual, not a smarter clinician.
A CHW who trusts the tool too far and skips escalation. A child whose case is borderline and gets a chatbot-generated reassurance instead of a clinic referral. We need hard rules about when the tool must say 'go to clinic, now.'
Restructure supervision: monthly group calls with a clinical officer. Better radios. We're doing both of these in parallel; they don't replace the reference need.
Praekelt's openCHW assistant. WHO's IMCI Digital Adaptation Kit. Dimagi's CommCare. Last Mile Health's Community Health Academy content.
Yes. Praekelt's openCHW is open-source, MoH-aligned in Ethiopia already, and has Tigrinya partial support. It is built for exactly our user. We almost missed it because we started in the wrong place (vendor pitches, not the global health commons).
Some dependency on Praekelt's release cadence. Mitigated by self-hosting. We'd contribute content back upstream.
Fork openCHW and add a hard-coded escalation rule: any query that mentions a child under 2 months or any red-flag symptom (convulsions, dehydration, difficulty breathing) returns 'Refer to clinic immediately' first, before any reference content.
Remove the 'differential diagnosis' module entirely. CHWs are not diagnosticians. Keep only protocol lookup, dosage tables, and red-flag escalation. Less is more faithful here.
Yes — the AI is strictly retrieval and reformatting. It does not generate clinical reasoning. It never says 'I think' or 'most likely.' The CHW remains the decision-maker; the tool is a manual she can talk to.
This team modeled the discipline of Reimagine well: they didn't dismiss the existing tool because it was imperfect, and they didn't rebuild from scratch because they could. They subtracted (removed diagnostic features), translated (Tigrinya, Amharic), and constrained the AI's role to retrieval, not judgment. The result is smaller, slower, and more accountable than what they originally wanted to build — which is usually the sign of faithful reimagining.
Start from this example as a draft. Every field will be pre-filled — edit freely. Your own context will surface as you go.
These examples are illustrative. Real discernments will be more complex, more painful, and more specific to your context. The library helps you see the shape of the work.
Examples paraphrased from common patterns observed in faith-based development AI projects. Not based on any single real organization.