A simulated clinic consultation: a doctor at his desk, an elderly patient and his son seated facing him.

Simulation · AI-generated people

Health · Language · Africa

Health care that understands you, in your own language.

BantuHealth is building AI that nurses and patients can trust in Bantu languages: a language assistant for nurses, a text-message version for any phone, and a safety test every language must pass before it is switched on.

Watch the clinic visit See the assistant Bring it to your city

Starting in Zambia English demo today Bemba · Chewa · Tonga · Swahili · Zulu: the goal Built with native speakers

The whole visit, from the knock to the see-off

Watch what we test.

A simulated clinic visit, start to finish: an older man with a painful knee arrives with his son, an AI plays the doctor, and the visit ends with a nurse and a goodbye. We record what the doctor was told, what it said and what it did, and check it against what really happened. This version is in English; the Bantu-language versions are the goal.

Simulation. The people and voices are AI-generated. Captions are burned in. Nothing here is a real patient, clinic or result.

Naming the body, in many languages

A second test: the doctor names the parts of a 3D body, one by one, in the singular and then the plural. Native speakers supply the words for every language. Here is the English version and the breadth of the data: 24 languages, 30 body parts, singular and plural. We pilot with the first select languages and scale to the rest with the same method. (The words themselves are kept out of this page on purpose: they are the answer keys of our test.)

Simulation. The body and the voice are AI-generated. Captions are burned in.

An elderly man leans on his adult son as they walk into a clinic room. — From our clinic simulation

The problem, in one clinic visit

A small misunderstanding can become a big one.

He walks in leaning on his son.
An older man with a painful knee, in a clinic that serves many languages.
The nurse and the patient don't share a language well.
The words are close, but not the same. Right or left? Once a day or twice? This week or next?
AI could help, if we know when it is safe.
Translation tools are arriving. Nobody has checked how they behave in Bemba, Chewa or Tonga. In health care, a wrong word costs more than no word.

Before an AI helps in a clinic, someone has to prove it understands the language.

What we are building

Three things that work together.

Each one is small enough to test properly, and useful on its own.

The nurse's language assistant

A nurse types or says an instruction. The assistant suggests how to say it in the patient's language and shows what it understood. The nurse always decides what is sent. It never diagnoses or prescribes.

Try the demo

A text-message version

Many patients have a basic phone and no data. The same help, by SMS: appointment reminders, clinic hours, and a plain-language reminder of what the nurse said, in the patient's own language.

See the text-message version

Safety tests for every language

We test AI in a simulated clinic visit (an AI plays the doctor) with native-speaker review. A language is switched on only when it passes. If it doesn't pass yet, it stays off.

See how we test

Try the nurse's assistant

See what a nurse would see.

Pick something a nurse might need to say. Then pick a language.

1. What does the nurse need to say?

2. In which language?

What it will never do

Diagnose, or suggest what illness someone has
Prescribe or change a dose
Send anything the nurse has not approved
Replace the nurse

BantuHealth · Nurse assistantdemo · scripted

Scripted demonstration in English. The Bantu-language versions are the goal and appear here only after native-speaker review. Illustration only; not medical advice.

How we test

An AI plays the doctor. We check what it understood.

This is a real simulation we built: an elderly man, his adult son and a doctor. It is the same visit every time, so different AI systems and different languages can be compared fairly.

The consultation room as the doctor sees it: an older man and his son seated in front of the desk. — Simulation · AI-generated people

Real situations. Greeting an elder, asking who came with him, which knee hurts, left and right from the patient's own side.
Scored by fixed rules, not by another AI. The same answer always gets the same score.
Checked first. We test the test: a perfect player, a random player and a silent player must score as they should.
Honest about what is unreviewed. Every score shows how much of the answer key a second native speaker has not yet reviewed.
No rankings of languages. The goal is to know where AI is safe to use, not to grade a language.

Ask for a live walkthrough

Languages

English today. Bantu languages are the point.

This site and the demo are in English for now, so that anyone can see how the idea works. The goal is every one of these in the patient's own language, built with native speakers.

How a language goes live

Collect

Native speakers give us the words for the body and for health, with consent.

Review

A second speaker and a clinician check them. We show what is still unreviewed.

Test

We run the AI through the simulated clinic visit in that language.

Switch on, or not

Only what passes is enabled, for nurses first. The rest stays off.

A 3D human body in which every body part is a separate coloured region with a label.

One body, every language.

In our 3D body every part has its own identity, such as the hand, the knee or the forearm. The same body and the same questions are used in every language, so results can be compared fairly.

Our vocabulary of the body and health already covers 24 Bantu languages, entered by native speakers. Second-speaker review is under way. We do not publish the words here until they have been reviewed.

Safe by design

Careful first. Clever second.

Health is not the place to move fast and fix later. These are the rules we build by.

The nurse is in charge

Every message is approved by a person. The assistant helps; it does not decide.

No diagnosis, no prescribing

It explains and translates what a nurse has already decided. Nothing more.

Tested before it is switched on

A language and a function are enabled one at a time, only after they pass.

Native speakers decide

Words are collected and reviewed by speakers of the language, not guessed by a machine.

Privacy by default

Testing uses actors and scripts, never patient records or recordings of real consultations.

Shared, so cities can trust it

What we build for a city's clinics is meant to be published openly so others can check it and reuse it.

For city leaders

Bring it to your city's clinics.

A first pilot is small, careful and led with your health team. This is what two years can look like.

Months 0–3

Set up

Agreement with the council, approvals, a local coordinator.

Months 2–8

Listen

Which languages do patients use? Where does language cause problems?

Months 6–14

Build & test

A prototype with nurses. Native-speaker review. Safety tests per language.

Months 12–21

Pilot

In a few clinics, with a nurse in charge, a log of every problem and a stop rule.

Months 20–24

A plain-language report, a checklist for the council, staff training.

What the city does

Names one contact person
Helps arrange access to a few clinics, with staff agreement and approvals
Takes part in reviewing results and decides how AI is used in its clinics
Gives a short, non-binding letter of intent to start

What the city gets

A clear report on where AI is, and is not, safe in local languages
A checklist for adopting or refusing AI tools in clinics
Training and paid local roles for coordinators and reviewers
Software and documentation published openly for other councils

Questions a mayor will ask

Straight answers.

What does it cost the city?

The city is asked for its time and access to a few clinics, not for its budget. We are seeking outside funding for the pilot itself. The first step is a short, non-binding letter of intent that commits the city to nothing.

Is this already used in clinics?

No, and we want to be clear about that. The clinic simulation and the safety tests are built and working. The nurse's assistant and the text-message service shown here are demonstrations of what a pilot would build, with your health team, and test carefully.

What if the AI gets it wrong?

A nurse approves every message before it is sent. The assistant never diagnoses or prescribes. A language is enabled only after it passes our tests, and the pilot keeps a log of every problem and has a stop rule agreed with the council.

Who sees patient information?

Testing uses actors and scripts, never patient records or recordings of real consultations. Any later use of real data would need consent, ethics approval and the council's agreement first.

Why Bantu languages, and why only English on this site?

Because that is where the patients are, and where AI is least tested. The site is in English so anyone can see how the idea works. The Bantu-language versions are the goal and appear only after native speakers have reviewed them.

What do you need from us first?

One contact person at the council, and a short letter of intent. After that: a conversation with your health officers about which languages your patients use and where language causes problems.

Get in touch

Let's talk about your city.

Tell us a little and we will reply with a short summary and a walkthrough of the simulation.