Loading...
Loading...
Loading...
Every Swahili reply is checked for noun-class agreement, verb form, financial terminology and register before it reaches you, and corrected where the check finds something it can fix.
We prepare the file. The lender interviews, verifies and decides.
These checks run on the assistant's own words, on the way to you, and on the catalogue the interface is built from.
The first four run on the reply itself before it is streamed. The rest are build-time gates: a Swahili key that goes missing does not quietly render English, it stops the release.
Swahili nouns fall into classes, and the adjective, the verb and the possessive all have to agree with the class of the noun they attach to. Machine output gets that wrong constantly, and a reader clocks it immediately: this thing was written for somebody else and handed to me second. For a product about money, in the language most of this market actually does business in, that is not a cosmetic problem. It is the reason people switch to English and lose half of what they meant.
Noun-class agreement, verb morphology, terminology and register are all examined on the generated text. Where the check can repair the sentence and the quality is below the floor, the repaired sentence is the one you receive.
Financial terms are held to a single canonical word each, so the same concept is not called three things across three screens. Terminology is corrected on the reply, not left to the model's mood.
When a word is not recognised, the assistant is biased toward asking you what you meant rather than assuming, and it is capped so a conversation does not turn into an interrogation. What you answer is kept.
Words are broken into morphemes and analysed against noun and verb roots, which is why a form the system has never seen can still be understood. Matching whole words would fail on the first prefix.
A key that exists in English but not Swahili fails the build, as does an empty value and a mismatched placeholder. There is no falling back to English, because falling back to English is exactly the mixing this is meant to prevent.
The language is resolved on the server before the page is sent, so a Swahili reader never sees an English flash while the page catches up. The assistant is held to the same rule: it understands you in any language and answers wholly in yours, including the greeting.
Swahili is not a feature that gets finished. Treating it as one is how products end up with a good first screen and a machine-translated fourth.
That is the test we hold ourselves to. Take it.