Skip to content

AI & Automation

Put AI to work on the tasks that eat your team's week.

We build agents, copilots and document pipelines on top of your own data — scoped to a measurable saving, evaluated before they ship, and governed so nothing leaks.

Book a discovery callBuild your plan
  • Dubai head office, GCC-wide delivery
  • Arabic and English throughout
  • Written scope and fixed timeline before you commit

You walk away with

  • RAG on your data
  • Workflow agents
  • Evaluation harness
  • Human in the loop

Typical timeline

6–14 weeks

Typical engagements start around

AED 30,000 per project

What good looks like

graded cases in the evaluation set
100–300graded cases in the evaluation set
process scoped per engagement, with a baseline
1process scoped per engagement, with a baseline
of answers traceable to a source document
100%of answers traceable to a source document

The problem this solves

Every board in the Gulf has asked for an AI strategy and most have received a chatbot nobody uses. The pilot demos well, then meets real documents, real Arabic, and real permissions, and quietly gets switched off. The failure is almost never the model — it is that nobody defined what success would look like, and nobody built a way to measure it.

Where this usually hurts

  • The pilot never became a product

    It demoed beautifully on curated examples, then met real documents, real Arabic and real permissions, and was quietly switched off.

  • Nobody defined success

    Without a measured baseline for handling time or error rate, the project can never be shown to have worked — so it gets cut in the next budget round.

  • It invents things

    A model answering from memory rather than from your documents will produce confident, plausible, wrong answers, and nobody can tell which ones.

  • Legal will not sign it off

    No logging, no PII handling, no record of what data grounded what answer — so the security review stops it before it reaches users.

We build AI systems that sit on your own data and do a specific job: answer questions from a policy library, extract fields from invoices and delivery notes, draft a first-pass response a human then approves, or route a case to the right desk. Every engagement starts by picking one process and agreeing what a saved hour is worth, because an AI project without a baseline can never be shown to have worked.

Retrieval before generation

A model that answers from memory will invent things. A model that answers from your documents, and cites which paragraph it used, can be checked. We build retrieval first — chunking, embeddings, a hybrid keyword-and-vector index, and permission filters applied at query time — so an answer is always traceable to a source the reader is allowed to see.

  • Answers cite the document and section they came from
  • Permissions enforced at retrieval, so nobody sees a file they could not open themselves
  • Arabic and English indexed together, so a question in one language finds evidence in the other
  • A refusal when the corpus does not contain the answer, rather than a confident guess

Evaluation is the deliverable

Before anything ships we build a test set from your real cases — usually 100 to 300 questions with expected answers written by whoever does the job today. That set becomes the gate: every prompt change, model upgrade and index rebuild is scored against it, so you find out that a change made things worse in CI rather than in a complaint.

100–300

graded cases in the evaluation set before launch

Humans stay in the loop where it matters

Automation earns trust by degrees. We start with the system drafting and a person approving, measure the edit rate, and only widen the autonomy where the edit rate says it is safe. High-consequence steps — anything touching payment, contracts or personal data — keep a named approver permanently.

Governance that survives an audit

Prompt and response logging with retention you control, PII redaction before anything leaves your boundary, model routing so sensitive workloads can run in-region or self-hosted, and a written record of what data trained or grounded what. This is what the UAE Personal Data Protection Law and your own security review will ask for, and retrofitting it is far more expensive than building it in.

What is included

  • Use-case scoping

    One process, a measured baseline, and an agreed number that says whether this worked.

  • Retrieval pipeline

    Ingestion, chunking, hybrid search and permission-aware filters over your document set.

  • Evaluation harness

    A graded test set from your real cases, run in CI on every change to prompts, models or index.

  • Human-in-the-loop UI

    Review, edit and approve screens, with the edit rate tracked as the signal for widening autonomy.

  • Guardrails and logging

    PII redaction, refusal behaviour, full prompt/response audit trail and retention you control.

  • Model strategy

    Routing between hosted and in-region or self-hosted models, with a written cost-per-task model.

What we build with

  • Models

    • Claude
    • OpenAI
    • Azure OpenAI
    • Open-weight models

  • Retrieval

    • pgvector
    • Hybrid search
    • Rerankers
    • Document parsing

  • Orchestration

    • LangGraph
    • Python
    • FastAPI
    • Evaluation harness

How engagements are sized

Three sizes, so the scope matches the problem rather than the budget matching a template. Every tier is a starting point — the number moves with what we find in discovery, and you see the revised figure before anything is signed.

  • Essential

    One clear problem, scoped tightly and shipped.

    From

    AED 30,000

    per project

    One well-defined task — document extraction, a policy assistant, a triage step.

    • Single use case
    • Retrieval over one corpus
    • Evaluation set
    Get a firm number

  • Growth

    The full engagement, with measurement and iteration built in.

    From

    AED 90,000

    per project

    A production system serving a team, with governance and measurement.

    • Multi-source retrieval
    • Human-in-the-loop workflow
    • Full audit logging
    Get a firm number

  • Enterprise

    Multi-entity, regulated, or integrated across several systems.

    From

    AED 240,000

    per project

    In-region or self-hosted models, multi-department rollout, regulated data.

    • Private model deployment
    • Multi-tenant permissions
    • Compliance documentation
    Get a firm number

Prices are in AED excluding 5% VAT. Other currencies on the pricing page.

How delivery actually runs

The phases below are what a typical engagement moves through, with the durations we plan against. You always know which phase you are in and what leaves it.

See the full process
  1. 1

    Scope and baseline

    1–2 weeks

    Pick one process, measure how it runs today — volume, handling time, error rate, rework — and agree the number that defines success.

  2. 2

    Retrieval and evaluation

    2–4 weeks

    Ingest and index your documents with permission filters, then build the graded test set from real cases before a single prompt is tuned.

  3. 3

    Build with a human in the loop

    3–6 weeks

    System drafts, a person approves, edit rate is tracked. Autonomy widens only where the numbers say it is safe.

  4. 4

    Roll out and measure

    2–4 weeks

    Live on a subset, the same metrics re-measured against the baseline, and a written decision on whether to widen, adjust or stop.

  • Scope and baseline · You get

    • Measured baseline
    • Scoped use case
    • Cost-per-task model

  • Retrieval and evaluation · You get

    • Permission-aware index
    • Graded evaluation set
    • Baseline scores

  • Build with a human in the loop · You get

    • Review and approve UI
    • Edit-rate tracking
    • Guardrails and logging

  • Roll out and measure · You get

    • Production rollout
    • Before-and-after report
    • Runbook and handover

Built for the UAE

What working with a Dubai partner actually changes

Most of what follows is invisible until it goes wrong — a supplier who cannot keep data in-country, an invoice the FTA rejects, a support rota that is asleep for half of your working day. These are the questions we answer before they become findings.

Two questions decide most UAE AI projects: can it work on Arabic documents, and can the data stay in-country. We test Arabic quality separately and report it as its own number, and we route sensitive workloads to models running in-region or in your own tenancy rather than assuming a US endpoint is acceptable.

  • Data residency, decided up front

    AWS me-central-1, Azure UAE North and OCI Dubai are all in scope. We map the whole data path — backups, logs, metrics and support access — rather than the primary database alone, and put the result in writing.

  • PDPL, and the free zones

    Federal Decree-Law No. 45 of 2021 governs most of the mainland; DIFC and ADGM entities fall under their own data protection laws instead. We establish which applies to your entity before we design anything that touches personal data.

  • VAT and e-invoicing that pass

    Five per cent VAT, FTA-compliant tax invoices, the audit file the authority can request, and readiness for e-invoicing as the mandate phases in. Configured in the build, not patched after the first filing is rejected.

  • Your hours, your languages

    Delivery runs on Gulf hours with real overlap with your team, in Arabic and English. Customer-facing interfaces, documents and training are produced in both where you need them, with RTL treated as a design requirement rather than a translation step.

Sectors we do this in most

Not the only sectors we work in — the ones where we have shipped this service enough times to know the regulatory edges and the usual traps.

  • Fintech

    Onboarding, KYC and dashboards that build trust and convert.

  • Healthcare

    Compliant, accessible experiences patients actually use.

  • Real Estate

    Listing and portal experiences that turn browsers into leads.

In-house, generalist agency, or us

An honest comparison, including where the other two options are the better call.

Hiring in-houseWorking with us
Time to first outputThree to five months to source, notice-period and onboard, in a market where senior specialists are scarce and expensive.Two to three weeks from signature, with people who have shipped this before.
Cost shapeFixed monthly cost plus visa, insurance, end-of-service and equipment, whether or not there is a full workload.Scoped to the work. Costs stop when the work does.
Breadth of skillOne or two specialisms per hire. Anything outside them gets improvised or outsourced anyway.Design, engineering, security and growth from the same team, without a handoff between vendors.
Institutional knowledgeStays in the building — which is the real advantage, and the reason to hire eventually.Documented and handed over. We write runbooks so you can take it back in-house.
When it is the wrong choiceRarely, once the workload is genuinely full-time and permanent.If you need someone in your standups every day for years, hire. We will say so.

Related work

Questions people ask before starting

No. We use enterprise API tiers where the provider contractually excludes your data from training, and for anything sensitive we route to a model running in your own tenancy or in-region. Which data goes where is written down before we build, not decided later.

Services that pair with this

Most engagements use two or three together.

Ready to talk about AI & Automation?

Tell us what you are building and where it is stuck. We will come back with a scope, a timeline and a number — usually within two working days.

Step 1 of 3 · What you need

Which services are you interested in?