Building
Kinyarwanda language
Text comprehension, code-switching with English and French, dialect awareness, formality registers, idiomatic and oral-tradition language.
Live partial inside Scribe → standalone in progress
A Kite initiative building purpose-built, sovereign AI for Rwandan health, education, agriculture, government, justice, finance, and culture. Small models. Real problems. Built in Kigali.
We can't build it all. We're building the layer young Rwandan engineers will build on top of.
Building, within Rwanda, the infrastructure the next generation of Rwandan builders will stand on. Language, voice, vision, reasoning — the substrate underneath every product Rwanda actually needs. The backlog below is a starting set; the real list is much longer, and grows with every builder who reaches out. Rwanda Scribe is our first piece, live in pilot.
A PhD takes a decade — sometimes longer. Plenty of Kinyarwanda AI papers have been written; the real problems on the ground remain unaddressed. Academic AI rewards novelty and benchmarks; it punishes the unglamorous work of integration, deployment, and maintenance — exactly the work Rwanda needs done.
That path is real, and we want people on it. Some Rwandans will rightly spend a decade in a lab — Rwanda needs them too. But not all of us. Not most of us. And meanwhile there are hundreds of 18-year-olds in Kigali, Musanze, Huye, Rusizi who could ship a product this year — a few have already shown up at our door.
We took the opposite trade. Solutions over citations. Production over publication. Practice over papers. And a layer that lets a kid in Nyabugogo build on top of it — without waiting ten years for a degree first.
Plenty of papers. Zero deployed clinics. The real problems remain unaddressed — and we're inviting Rwandan engineers, at any stage of any path, to take the opposite trade with us.
Build alongside credential, not because of it.Twelve kinds of understanding that combine to make a Rwandan AI stack. Each is at a different stage. None of them are products on their own — they're the substrate that products are built from. Hover any block to see which Layer Two products fuse it.
Text comprehension, code-switching with English and French, dialect awareness, formality registers, idiomatic and oral-tradition language.
Speech recognition, multilingual TTS with code-switching, speaker separation, audio events, environmental sound understanding.
Document OCR for Rwandan handwriting and prescriptions, photo-based pest and disease ID, medical imaging, traffic and crowd vision.
Rwandan proverbs, social conventions, Imihigo culture, kinship and respect registers, oral tradition, family and community context.
Rwandan law codes, RDB regulations, RRA tax code, Mutuelle and RSSB rules, labour and land law, sector-specific compliance.
Clinical, agricultural (RAB), educational (CBC), financial — sector-specific knowledge layers that vertical agents reason over.
Multi-step reasoning, contextual judgment, Rwandan-aware planning — how an agent decides what to do next in a real workflow.
Long-term user, family, and organisational context. Remembers your village, your crop, your cooperative, your last appointment.
Specialist agents collaborating on one task — the clinic agent asks the regulatory agent that consults the clinical agent.
Models that run in Kigali, on-prem, or at the edge. Offline-capable for clinics and co-ops where internet cannot be assumed.
Kinyarwanda harm detection, misinformation defence, bias evaluation in low-resource settings — evals in Ki, not translated.
Anonymisation, consent flows, in-country data residency, audit trail. Customer data trains the customer's model, not someone else's.
The frontier isn't one model doing one task. It's multiple kinds of understanding fusing into one experience that meets a Rwandan in their context. Five examples of where we think the line goes. All concept-stage.
One agent with four senses, working alongside the nurse.
ConceptWatches the consultation camera, hears the Kinyarwanda conversation, reads the patient chart, drafts the clinical note, and flags the drug interaction — all in real time, on a tablet next to the nurse. The nurse stays in charge; the agent does the typing.
A unified brain for Kigali's emergency response.
ConceptKigali already runs dense CCTV. Add Kinyarwanda audio understanding, GPS, weather, and traffic feeds — and one agent can route the right responder to the right incident in seconds, with the dispatcher staying in command.
One Kinyarwanda voice that knows your whole family's context.
ConceptKnows your Mutuelle, your Irembo applications, your MoMo flow, your kids' school calendar, your cooperative meetings. Answers in Ki. Remembers what mattered last month. Coordinates appointments without you opening five apps.
The co-op chairperson's full back-office, voice-first.
ConceptOne Kinyarwanda voice that runs the co-op's pricing line, member ledger, RAB advisories, weather feed, and SACCO loan applications. The chairperson speaks to it; members call into it; the records survive a phone change.
For traders moving goods between Rwanda, DRC, Uganda and Tanzania.
ConceptKnows each border's current paperwork, exchange-rate windows, and customs quirks. Drafts the documents in the right language, in the right format, and stays current as rules change. The trader's accountant in their pocket.
We can't build all of this. We're not trying to.
Rwanda has thousands of kids who could ship a product today. Many will never reach a university — not for lack of talent, but for lack of means. Many who do reach one will spend years on a degree that doesn't help them ship.
We're building Layer One — the language, voice, vision, regulatory, and reasoning substrate — so an 18-year-old in Nyabugogo with a phone and an idea can ship a Mutuelle navigator, an MoMo fraud detector, or a Kinyarwanda tutor.
Without the degree. Without the lab. Without the gatekeepers.
And we don't see you as integrators. The tools available today let a determined 18-year-old train models, run experiments, and find the doctors, lawyers, or farmers willing to validate what they build — work that a decade ago needed a PhD.
Technology bridges the gap. We bridge the rest: scaffolding, conversations, validation paths where we can — and the right introductions, as we build the relationships to make them.
Rwanda needs engineers more than it needs degrees.
Talent doesn't need permission. It needs scaffolding.
We don't run a packaged accelerator. Our resources are limited and we allocate them deliberately. We'd rather back ten kids who ship than a thousand who never start.
Tell us what you'd build and what you've already made. Show a prototype, a repo, a video, a clear plan — whatever proves you can ship. No degree, no school, no credential required.
Layer One access. A specific capability. Deployment help. A door we might know how to knock on — or a sense of who to approach if we don't. Be specific.
If we see real drive, real talent, and a real plan, we work out together how we can support you. The conversation is the test. Terms come later, in private.
If your build needs a Layer One capability we haven't shipped yet — a dialect, a regulatory corpus, a deployment path — and your case is real, we can prioritise it. Demand from real builders moves our build order forward.
The backlog is below. Pick one. Or bring something we haven't thought of.
Apply to build on Layer One →A growing backlog of real Rwandan problems. We're not building them all ourselves — we can't, and we shouldn't. Each card describes the problem honestly, the building blocks it would require, and how close to a pilot it is.
Capabilities we are putting into the hands of partners. One is live today. The rest are in active build.
Production-grade accuracy on Kinyarwanda where generic models hallucinate. Live inside Scribe. Standalone productisation in progress.
Kinyarwanda, English, and French in a single voice. Code-switches mid-sentence the way Rwandans speak. Voice cloning, natural prosody.
Translation across Ki/Eng/Fr that handles the mid-sentence switches and idioms generic translators flatten. Early voice prototype runs Ki→Eng/Fr today — quality not yet usable.
Knows who is speaking and tracks the conversation across turns. Live inside Scribe.
Models that run in Kigali, on-prem, or at the edge. No data leaves the country unless the customer chooses.
Kinyarwanda-first audio & video transcription with speaker labels, timestamps, and code-switch detection. scribe.rwandaai.work →
Our position, written down so partners and builders know what to expect — and so we have something to be held to.
One model per problem, fine-tuned on Rwandan data — not a generic LLM bent toward Kinyarwanda.
Customer data stays in Rwanda or on-prem. Your data trains your model, not someone else's.
Anything we build must work without a keyboard. Rwanda is a voice-first country.
Hardware Rwandan institutions can actually afford and own. Not someone else's GPU cluster.
Publish where it grows the ecosystem. Hold what creates customer value.
Today, every Scribe correction enters a retraining dataset. We haven't run that retraining yet — the data is the seed, the next model is the payoff. That compound is the moat.
We're building with institutions and with builders simultaneously. Both move us forward. Neither gets a separate site.
If you can ship, we want to know.
Where deep institutional context matters.
We're choosing the first pilots now. The partners we work with first shape what gets built. We read every message. Replies usually within 48 hours.