Umuseke — Kinyarwanda language AI

Four things that break when you use a general model here.

Each one has a specific engineering answer. None of them is “use a bigger model”.

What happens between a voice note and an answer.

Four stages, each of which you can inspect and log.

Four deployment shapes we already support.

Each one differs mainly in how the answer gets back to the person who asked.

The playground runs the same engine.

Switch languages, adjust the decoding temperature, inspect the entity parse and copy the request as code. No signup, no email gate — if it does badly on your text, we would rather you found out here.

Reasonable objections.

Umuseke

General models stumble on agglutinative, dialect-rich Bantu languages — they fragment the words and invent the context. Umuseke is a language engine trained for Kinyarwanda morphology, grounded in local agricultural and public-sector sources, and packaged to run where your data already is.

Acoustic & dialect ingestion

Speech captured in the field is normalised for pitch accent and regional variation before anything else happens.

Morpho-syntactic engine

Words are decomposed into their grammatical parts, so meaning survives the agglutination that breaks generic models.

Grounded generation

Answers are drawn from a vetted corpus, with the source retrievable rather than implied.

Sovereign dispatch

Results return over the channel that suits the user — app, SMS, IVR or API — without leaving your infrastructure.

A byte-pair tokenizer calibrated on 16 Bantu noun classes and infix grammar, so a word survives tokenization intact.

One Kinyarwanda word can carry what English spreads over a clause. General tokenizers shatter it into fragments that mean nothing.

Retrieval grounded in agricultural extension handbooks and national gazettes, with a citation attached to every answer.

Frontier models improvise agricultural advice for a context they have never seen, confidently and incorrectly.

On-premises appliances in Kigali or Montreal, and a fully offline mode for field devices.

Sending citizen or farm records to a third-country API is often not legal, and rarely wise.

4-bit quantized weights running on rugged quad-core field tablets, with no network dependency.

A round trip to a distant data centre is not viable on an intermittent rural connection.

Agronomic voice advisory

Farmers ask a question by voice in Kinyarwanda and receive a grounded answer. Works over IVR and messaging gateways, so no smartphone is required.

Government & digital public infrastructure

Translation and document understanding for public services, deployed inside the institution rather than called out to a foreign API.

Financial & insurance verification

Claims and loan documentation read in the language they were written in, cutting the manual translation step out of the approval path.

Sovereign knowledge bases

Internal documentation made searchable across Kinyarwanda, French and English, hosted in private clusters in Montreal and Kigali.

Runs where the data is

On-premises or fully offline, so field recordings never leave the country to be understood.

Tokenized for Bantu morphology

A word survives tokenization intact instead of shattering into meaningless fragments.

Grounded, with citations

Answers are drawn from a vetted corpus and carry the source they came from.