Generic model
The gap
Nobody ever wrote this language down.
Moroccans do not type Darija in Arabic script. They type Arabizi, in Latin letters mixed with numbers. Almost nothing on the open web looks like that, so nothing inside a frontier model does either. Ask one for a bank rule in Darija and it answers in confident, fluent Standard Arabic that nobody actually speaks.
And here, the data that matters cannot legally leave the country. Dialect and sovereignty are not two problems. They are the same wall, and it is why this market is still empty.
safiDarijal'connexionFrenchtayraDarija · flips with contextdabaDarija · now
Four words, two languages, and one that changes meaning with the sentence around it. This is what a real customer message looks like, and no public dataset contains it.
The stack
One stack. The whole conversation.
Three layers, shipped in order, each one paying for the next. Every layer runs on the client's own hardware.
01
Understand
A Darija-native language model. One core, one hot-swappable adapter per industry, fine-tuned on data we validated ourselves.
Live · deployed on-premise02
Listen
Speech to text for real phone calls, not studio audio. Call archives first, then real-time voice.
In training03
Speak
A natural dialect voice that closes the loop, from a spoken question to a spoken answer.
NextWho it is for
Built for the places where the data cannot leave.
Regulated industries with millions of customers who speak dialect, and a legal obligation to keep every word at home.
Pilot in discussion
Years of recorded calls nobody has ever listened to.
- Transcribe the archive in dialect, on your own servers.
- Find the calls that matter: billing disputes, outages, customers about to leave.
- Turn a dead archive into a monthly picture of why people actually call.
The moat
We do not collect the data. We create it.
Public dialect datasets are written in Arabic script by journalists. Useless for a real phone call. So we built our own supply chain instead of renting someone else's, and we run it every day.
Our own factory
Raw material in, training material out. Nothing we ship depends on a public dataset that could disappear or was never right in the first place.
Our own annotation platform
annotate.lessan.co, where native speakers validate everything a model is allowed to learn, with quality control above them.
Measured, never assumed
Every batch is measured against a test set we froze on day one. Nothing ships on a feeling. We walk clients through the curve under NDA.
Infrastructure, not a wrapper. What we sell is not a call to somebody else's API with a prompt in front of it. It is a model we trained, on data we own, running on your hardware.
annotate.lessan.co, our annotation platformSovereignty
Air-gapped by design. Not by promise.
No API call. No telemetry. No anonymised exception. The model runs on a single GPU inside your datacenter, and your customers' words never cross your own perimeter. Not for inference, not for training, not ever.
Built for Law 09-08 environmentsZero external callsOne GPU
1 × L40-CLASS GPURUNNING
api.openai.comus-east-1.awstelemetry.servicecloud backup
The architecture
One core. Every industry. Any dialect.
A dialect core with hot-swappable adapters per industry. Replace Darija with Wolof, Tagalog or Javanese and nothing in the architecture changes. Only the people who validate the data do.
core · darijawoloftagalogjavanesepidgindarija
Live today
masignal.pro
Our public product for the Casablanca Stock Exchange. It follows five investor groups, five thousand people, and turns their messages into a readable signal: what the market feels, stock by stock, day by day. Written in Arabizi, read by our stack. Free and open to anyone.
It is also our proving ground. Every product we run teaches the factory what real Darija looks like, and sets the bar our own models have to beat before they take over.
masignal.promasignal · masi intelligence
3tini l'analyse dial IAM
IAM · 112,40 MAD ▲ +1,8%
RSI 58 · support 109,80
Volume ×1,6 vs 30d average
RSI 58 · support 109,80
Volume ×1,6 vs 30d average
Where this goes
Morocco is day one.
Forty million people here. Two billion across MENA, Africa and South East Asia living under the same two constraints: a dialect no model learned, and a law that keeps their data at home. The factory is software. Opening a new dialect means recruiting native speakers, not rebuilding anything.
Morocco liveEgyptSenegalNigeriaIndonesiaPhilippinesand every market like them
Not a launch calendar. A map of the same wall: a dialect no model learned, and a law that keeps the data at home.
Shipped
Darija language model, on-premise
Our own fine-tuned model running on a client's servers, not ours.
Now
Speech to text for real calls
Validated hours going in every week, error rate measured against a fixed test set every time.
Next
Dialect voice
The third layer, so an agent can answer out loud in the language the customer used.
Then
A second dialect
Same factory, same architecture, a new pool of native speakers.
See it run inside your walls.
Bring your hardest recordings and your compliance rules. We will show you what runs today, and what does not yet. We open a few pilots at a time, and we choose them carefully.
Built by two engineers who bootstrapped their previous company to €1M a year, then automated it. We ship, then we talk.
45 minutes · your infrastructure, your questions, our founders