Fluent Is Not the Same as Understanding

When I first started thinking about ATAS, my plan was simple: build an AI model that understands Kinyarwanda. At the time, the big AI systems could barely handle the language, and it felt obvious that someone from here should fix that.

Then something unexpected happened. The models kept getting better. Month after month, they wrote Kinyarwanda that Rwandans could read without wincing. The problem I had set out to solve was being solved by someone else, faster than I could have done it alone.

That could have ended the idea. Instead, it made me look closer at what was still broken.

Speaking a language is not knowing a place

My father was one of the first people to show me the problem, without meaning to. Even when language was not the obstacle, he struggled to get real value from these tools. The words were fine. What was missing was an understanding of how things actually work here.

Take a simple, hypothetical example. Someone asks an AI how to start a small business. The answer comes back organized and confident: register the business, open a bank account, write a plan, find your customers. Every step is reasonable. But it is the answer you would give to anyone, anywhere. It doesn't know which office you would really have to visit, how the local market is organized, or what a person in a small town can and cannot reach.

A place existing on a map does not mean an AI understands it. Knowing that a location exists is very different from knowing what it means to live and work there.

The problems that never show up in a demo

Once I started looking for this pattern, I found it everywhere.

A citizen needs a document such as a birth certificate, and the process still often starts with a trip to the local administration office at the umudugudu level, for something that should not require moving at all. A patient needs help from a doctor quickly, and there is no fast way to reach one. A person calls a service center and waits two hours for someone to answer.

None of these are language problems. Translating them into Kinyarwanda fixes nothing. They are problems about how systems work: who is responsible for what, which steps really happen, what people can and cannot reach. And they show up across healthcare, agriculture, business, government, and ordinary daily life.

Why the big labs will not fix this for us

I don't say this as criticism. The major AI labs are building general intelligence: systems meant to be useful everywhere. That is a huge and valuable goal, and it is also the reason they will not spend their effort on the local details of one country's systems. Generality and local context pull in opposite directions.

Someone has to build the layer in between, a form of intelligence that knows a specific place well enough to be useful in it. I call that contextual intelligence, and IMIZI is ATAS's long-term attempt to build the infrastructure for it.

I want to be clear about where that stands. IMIZI is a research direction, not a finished product. I have a clear picture of the problem and a direction for the answer. I have not proven it yet.

What I have that the big labs do not

If I have an advantage here, it is not resources, money, or a large team. It is simple: I am a citizen of the place I am trying to build for. I speak the language natively. I know what these systems feel like from the inside. And I can collect the kind of data that barely exists on the internet, because it lives in daily life here and not in documents anyone ever posted online.

That is a narrow advantage, but it is real, and it points to where the work has to start: with the people and realities the technology is meant to serve, before the model.

The question I keep asking

Every time I look at a new AI product, I ask the same thing: does it understand the language, or does it understand the person?

Language is the doorway. Context is the room. Rwanda deserves AI that knows the difference.