Custom AI development · UAE
A build has to beat a hosted call at something. Here it beats it in two places: a labelled set in a language the general models barely saw, and a provable processing boundary.
Tell us what data you holdWhat custom AI development is
Most of what is sold as AI development is a prompt on somebody else’s model. A build has to be better than that at something specific, and the something is usually your data.
Your documents, your history, your images, labelled. The general models were pretrained on very little Gulf Arabic, so a labelled set drawn from your own bilingual paperwork is an advantage no subscription can sell you.
Where the model actually processes the data, written down and demonstrable. There is no general rule here that the data must stay in the country, but there is a transfer question, and it is answered differently depending on where you are registered.
The weights, the training code, the retrieval index and the runbook, in your name. A model you cannot retrain is a subscription with a longer contract, and the difference only becomes visible the month you want to change something.
What we build
Six things get commissioned under one name. Which one you need is decided by the data you already hold, not by the outcome you want.
A model that answers from your contracts, policies and filings rather than from what it remembers, with the source shown beside every answer
answers a person can check against the page they came from
The documents this market actually runs on, read and extracted: a trade licence, a tax invoice with a registration number, an identity document, a tenancy contract, a bill of lading
fields, not a summary
A model taught your vocabulary and your register, which is the case worth funding when the general one keeps getting your terminology wrong
a model you hold the weights to
Demand, risk or failure predicted from records nobody else has, evaluated against a held-out period rather than against the months it was trained on
a number with an error bar on it
Defects, counts or readings from cameras you already run, which is a labelling engagement before it is a modelling one
a labelled set that keeps its value after the model is replaced
A model given tools and a narrow remit, with every action logged and reversible
a bounded agent rather than an open-ended one, because the open-ended ones are the demos
Fine-tune it, or retrieve over it
Both put your data behind the model. They differ in what they need, what they fix, and what happens the week the document changes.
| Fine-tuningTeach the model your material | RetrievalLet it read your material | |
|---|---|---|
| What it needs | A labelled set, large enough and consistent enough to teach from. | The documents you already have, indexed. No labelling to begin with. |
| What it fixes | Vocabulary, register and format. How the model talks about your work. | Knowledge. What the model knows about your work, and where it read it. |
| When the source changes | A retraining run, on a schedule somebody has to own. | A re-index, which is a job that runs overnight and needs nobody. |
| What neither fixes | Data that is wrong, missing or contradicts itself across systems. | The same. Both inherit whatever the source documents already say. |
| Where the data goes | Into the training run, wherever that runs, and then into the weights. | Into an index you host, and into each call as context you can log. |
Retrieval, for most buyers, most of the time. It answers the question people actually have, which is what the organisation knows rather than how it phrases things, and it survives the documents changing. Fine-tuning earns its cost when the language itself is the problem: a register, a vocabulary or a dialect the general model handles badly, which in this market is a real and specific case. Often the answer is both, in that order. If what you need is a conversational product rather than a model, that is argued separately.
The UAE context
Webzenia has worked with Gulf clients since 2018. Three observations, and the first two are the business case for building anything.
Research on Arabic language and speech models finds the dialects diverge from Modern Standard Arabic in vocabulary, form and structure, and that a shortage of annotated Gulf material is among the largest barriers to accuracy. A Jebel Ali freight forwarder under JAFZA holds exactly the annotated material that is missing.
Transfer is permitted where the destination offers adequate protection, and otherwise on safeguards, consent or contractual necessity. No adequacy list has been published, so nothing is cleared by reference. A DIFC-registered fund sits outside the federal regime entirely, and mainland is not adequate to it either.
A document model taught on today’s invoice formats has a known shelf life for that one document type, and the plan should say so rather than discovering it. A Sharjah SAIF Zone manufacturer with years of scanned paperwork is training on a format with an end date attached.
What the engagement covers
Six stages, and the first decides whether the other five are worth starting. Most of what goes wrong in an AI project went wrong in the data.
What you hold, how much of it is usable, how it splits by language, and what has to be labelled before anything can be trained
a readiness read, and sometimes the verdict that there is not enough
The runs, the hyperparameters and the decisions behind them, recorded so a result can be reproduced rather than remembered
a model and the record of how it came to exist
A held-out set the model never saw and a threshold agreed before training rather than after it
a pass or a fail, and no argument about what good looked like
The model served where your registration requires it to run, with the path the data takes documented rather than described
an architecture diagram a compliance function can read
Live performance tracked against the evaluation baseline, with a threshold that raises rather than a chart nobody opens
notice that the model is getting worse before a user tells you
Weights, training code, the evaluation suite, the model card and a runbook, in your accounts
a model your own team could retrain without us, which is the test of whether you own it
Our stack
Chosen for what each does to the boundary rather than for what it can do. Select one to see where your data ends up.
Training and fine-tuning happen here, and the artefact that comes out is a set of weights you hold rather than an account you keep paying for.
We record every run and the decisions behind it, so a result can be reproduced a year later by somebody who was not there. A model nobody can retrain is a subscription with extra steps.
How the build runs
The first phase can end the project, and often should. Reading the data first is what stops a model being trained on data that could never support one.
We establish what you hold, how much of it is usable, and how it divides between Arabic and English, because the second number decides whether the language advantage is real. In parallel we read where the entity is registered and what that permits, since a mainland company, a DIFC entity and an ADGM one are answering different questions. Both reads happen before anything is trained.
The threshold is set before training rather than after, and the model is measured on data it has never seen. The hosted model is benchmarked in the same exercise, as the baseline the build has to beat: where it wins, the honest outcome is that there is no project. Every run and the reasoning behind it is recorded so the result can be reproduced by somebody who was not in the room.
The model is served in your own account in a region chosen against your registration, with the data path written down rather than described. Live performance is then tracked against the evaluation baseline, because a model does not fail loudly: it gets slowly worse as the world it was trained on moves, and the first person to notice is usually a customer.
Documented against your registration, so where it processes is a diagram rather than an assurance.
Our commitment
An AI build fails at the data and at the boundary, months before anybody looks at the model. These four are in the scope for that reason.
We agree the target before training.
What good looks like is written down while it can still be argued about, and the model is measured on data it never saw. A result reported against a threshold chosen afterwards has been marked by the person who sat the exam.
Where it runs is documented.
Where the model runs, what leaves it and under what basis, written as a diagram your compliance function can read. Where your registration makes a transfer a question, we answer it before the architecture rather than after the audit.
The model and the code are yours.
Including the labelled set, which outlives every model trained on it and is the part that actually appreciates. The test of ownership is whether your own team could retrain it without us, and we hand over enough that they could.
We tell you when to buy instead.
Often it does, and saying so ends the engagement at a data read rather than at a disappointing launch. A firm that sells custom models and never recommends against one is not giving you an assessment.
Common questions
Keep exploring
What your licence and data regime allow, ranked and costed.
Answers from your documents, in the Arabic customers write.
Agents that answer your published line, in Arabic and English.
Outbound calling built inside the telemarketing rules.
Threads that finish in writing, in Arabic and English.
Rewrite the process for this market, then automate it.
Wire the two systems either side of the hand-off.
A CRM that knows which of your companies is selling.
The suite configured around where data sits and how invoices leave.
Next step
Describe the data you hold and where the entity is registered. We will say what could be built on it, where it would have to run, and whether a hosted model already does the job.
Tell us what you need.