A mechanic’s diagnosis may move between English and Twi because each language carries a different part of the problem. Nkomo is designed to follow that code-switched explanation as one conversation, so the driver can describe the warning light, imitate the sound, and continue the diagnosis without choosing a language again.
In 1983, Air Canada Flight 143 was in the air when its two engines lost power. The Boeing 767 had taken off with far less fuel than the crew believed was aboard.
The aircraft’s fuel system measured mass in kilograms. During refuelling, figures involving pounds and kilograms were confused. The numbers looked usable, but their meaning had changed along the way.
Captain Robert Pearson and First Officer Maurice Quintal had to bring the powerless aircraft down. Quintal suggested the former Royal Canadian Air Force station at Gimli, Manitoba. Pearson glided the aircraft there and landed it without engine power. The incident became known as the Gimli Glider.
The Canadian Aviation Safety Board documented the chain of events in its investigation report. The lasting lesson is uncomfortable because the individual numbers were not meaningless. The failure happened in the conversion between systems.
One diagnosis can contain two languages
A driver in Kumasi might begin with the dashboard:
“The engine light came on after I started the car.”
Then the explanation changes shape:
“Na sɛ mehyɛ accelerator no a, ɛyɛ sɛ krrr-krrr wɔ front ha.”
The English identifies a warning. The Twi places the sound inside an action and a location: it happens when the accelerator is pressed, and it seems to come from the front. The “krrr-krrr” matters too. It is an attempt to reproduce evidence that may be difficult to describe with a formal mechanical term.
That whole statement is one diagnostic thought.
A system that treats the English and Twi as separate requests can lose the relationship between the dashboard light, the accelerator, and the sound. It may answer only the last sentence. It may ask the driver to translate. It may behave as if the conversation restarted when the language changed.
Nkomo lets the driver type or speak that explanation in natural Twi, Ghanaian English, or a mix of both. In voice mode, the driver can hold to speak, hear the response, and interrupt naturally when a detail needs correcting.
The point is continuity. The warning light and the sound belong to the same problem.
This is also why a language selector can create more friction than its label suggests. What Happens When a Language Selector Interrupts a Bilingual Thought? looks at what gets lost when a person must stop and classify their speech before continuing it.
Test the reasoning, not the vocabulary list
A useful test should go beyond asking whether an assistant knows the Twi word for engine, brake, or battery. Vocabulary recognition cannot show whether the assistant followed the diagnosis.
Give it a connected sequence:
“The warning came on this morning. Sɛ megyina traffic mu a, engine no wosow kakra. But when I move, it settles.”
Then ask a follow-up without repeating everything:
“So which part should I mention first to the mechanic?”
A strong response should preserve the structure of the account. The warning appeared in the morning. The shaking happens while the car is stationary. It settles after the car moves. The assistant should help organise those observations without pretending to know the mechanical cause.
That last boundary matters. A conversational companion can help someone prepare a clear description, identify missing details, or turn scattered observations into a useful note. It should not manufacture certainty about a fault that requires inspection.
Broad claims about multilingual ability are weak evidence here. Current Nsanku research evaluates zero-shot translation between English and 43 Ghanaian languages using 300 sentence pairs for each language. Its results vary substantially across proprietary and open-weight models. That variation reinforces a practical rule: test the language, the task, and the switching pattern people will actually use.
For Nkomo, the meaningful question is not “Does it support multilingual conversation?” Ask whether it follows this specific driver from English warning light, to Twi sound description, to English follow-up without dropping the thread.
Privacy belongs inside the test
A car diagnosis may include more than engine symptoms. A driver might mention where the car is parked, who usually drives it, a recent journey, or a voice note intended for a particular mechanic.
Nkomo makes the storage choice explicit. Cloud consent can be set to never, ask each time, or this session. Conversation history stays on the device under the user’s control. Turning history off purges it immediately. Data export and account deletion each take one tap.
Those controls should be tested with the language behaviour, not considered later. Speak the mixed-language description. Check whether the conversation follows the thought. Then check what was sent, what remained on the device, and whether deletion works as expected.
Errors should also be visible. If part of the request fails, Nkomo shows the error instead of swallowing it. Silence can look like understanding, which makes hidden failure especially risky in a diagnostic conversation.
Air Canada Flight 143 landed safely at Gimli, but the incident remains memorable because a conversion changed the meaning of the fuel figures. A bilingual car explanation carries a smaller consequence, yet the communication problem has the same shape: every part may appear understandable while the relationship between them disappears.
Test Nkomo with one real explanation you would give a mechanic. Begin in the language that comes naturally, switch when the sound or symptom is easier to express another way, then ask a follow-up that depends on both parts. The reply should carry the whole diagnosis forward.
Comments
No comments yet.