Why We Built a Swahili-Language Tutor into Insight

Ubunifu Insight includes an Education Tutor that teaches in Swahili. We built it because language is not a cosmetic setting in a learning conversation. It affects how a learner asks a question, interprets an explanation, and decides whether to ask for clarification.

This is a focused product decision, not a claim that one AI agent solves access to education. A tutor response can be useful and still be incomplete or wrong. Curriculum, teaching practice, connectivity, device access, and human support all matter beyond the chat window.

Within those limits, the language of the interaction is still worth designing carefully.

Translation is only one layer

A translated menu can help someone navigate a product. It does not ensure that the main interaction works naturally in the same language.

For a tutor, the main interaction is the explanation itself. A learner should be able to phrase a question in Swahili, receive a response in Swahili, and continue the exchange without first rewriting the problem in English. Follow-up questions matter because understanding rarely arrives in one answer.

This also creates harder product questions. Some school terms may be taught in more than one language. Mathematical notation should not change when the surrounding explanation changes. A technically correct translation may still sound unnatural or use vocabulary that is wrong for the learner’s level.

Those are interaction and evaluation problems, not simply interface localisation.

Familiar examples can reduce unnecessary decoding

One demonstration in the Tutor explains fractions with a chapati divided into equal parts. The mathematics is ordinary: one whole divided into four equal pieces gives four quarters. The value of the example is that the learner can focus on the fraction rather than first interpreting an unfamiliar reference.

Local context should not become decoration or stereotype. Not every explanation needs a Tanzanian object, and no single example represents every learner. The example is useful when it makes the concept more concrete. If it distracts from the concept, it has failed.

That suggests a practical standard: choose an example for instructional clarity, then check both the subject matter and the language.

A specialised agent creates a clearer contract

Insight is a document AI product with several specialised agents. Giving the education use case a distinct Tutor helps define what kind of response is expected. The goal is explanation and guided understanding, not simply producing an answer as quickly as possible.

A clear role also makes evaluation more specific. We can ask whether the response:

• addresses the learner’s actual question;
• uses understandable Swahili for the intended level;
• preserves correct notation and subject facts;
• explains the reasoning rather than only stating a result;
• uses an example that supports the concept; and
• acknowledges uncertainty instead of inventing a fact.

These checks are more useful than saying an agent is “localised” and leaving the term undefined.

AI is support, not the authority

Language fluency can make an incorrect answer sound more trustworthy. That is a central risk in an AI tutor, not a small disclaimer.

Learners should be able to compare important explanations with course materials and teachers. Educators need a way to review how the tool handles curriculum topics, ambiguous questions, and age-appropriate language. Product teams need test sets written and assessed by people who understand both the subject and the Swahili being used.

We should also distinguish a convincing demonstration from broad evidence. A strong fractions answer shows that a particular interaction works. It does not establish accuracy across mathematics, science, history, or every curriculum level. That requires repeated evaluation.

Why this belongs in Insight

Insight already works with documents, structured extraction, templates, and specialised AI agents. The Tutor explores another form of document-grounded work: helping a learner engage with educational material through a language they can use directly.

The broader product principle is not “everything must be in Swahili.” Tanzanian classrooms and workplaces are multilingual, and people move between languages depending on the task. The principle is that English should not be an unexamined requirement when the interaction can be designed responsibly in Swahili.

We built the Tutor to make that principle concrete. Its quality should be judged by the explanations it gives, the limits it communicates, and the evaluation behind it, not by the novelty of seeing Swahili in an AI interface.