Comprehension Over Translation

The gap between translation and meaning

Millions of Kenyans deal with government forms printed only in English, written in legal jargon. For someone more comfortable in Kiswahili or Gikuyu, one misunderstood field can mean a rejected application and a wasted trip to the Huduma Centre: lost time, lost money, and another day off work.

The obvious fix is translation, and it doesn’t work. Swapping labels word for word tells you what a field is called, not what the registrar actually wants in that box. The confusion lives in meaning, not language.

So Raia, “citizen” in Swahili, is not a translation tool. It’s a comprehension tool. The idea came out of an ongoing collaboration with a friend who runs the Swahili Verse Project; they bring the language depth and on-the-ground perspective, and I design and build the product.

One field at a time

The hardest design decision was refusing to show the form.

Every form tool defaults to the whole document: all the fields, all at once, because that’s what the paper looks like. But the whole document is the problem. It’s the wall of English legalese that made someone need help in the first place. Raia shows one field at a time instead: what this field is asking, why the government wants it, and a place to answer, in your language. The trade is real, since you give up the at-a-glance overview that confident users want. For the person Raia is built for, comprehension beats overview, and the intimidating document becomes a conversation.

Answering is voice-first: speak your answer in Kiswahili, Raia confirms what it heard, and you move on. Field explanations can be read aloud too. Voice input works in all three languages; text-to-speech only where a usable voice existed, which Gikuyu didn’t yet have when I built it. Designing inside that constraint honestly beat pretending it wasn’t there.

If the confirm-before-act pattern sounds familiar, it’s the same trust design as Clara, at higher stakes: there the cost of a bad parse was a miscounted vacation day; here it’s someone’s national identity card application.

The person owns the artifact

Raia never files anything on anyone’s behalf. At the end, it maps your answers back onto the official form layout as a PDF you download, or emails you your own summary. The AI explains and guides; the citizen sees, confirms, and owns the actual document. For a civic tool, I think that line - the AI advises, the person acts - is not a limitation. It’s the design.

Privacy by architecture

For a tool handling identity documents, “we don’t store your data” is a promise. Raia makes it a structural fact instead: there are no accounts, and there is no database. Answers live in memory and vanish when the tab closes. Voice recordings go out for transcription and are immediately discarded. Data leaves the device only when the user explicitly exports it. You cannot leak what you never stored.

Where it stands

The National ID application (ND1) is live, with birth certificates and business permits next. The first version was built in a weekend with Claude Code, working from the official form scans. A few people in Kenya have used it, and that’s the honest status: shipped, early, and finding its first users.

Cover photo: Nairobi, by Michael Njoroge on Unsplash.