All services

AI & automation

AI implementation, measured before it ships

We connect language models to the data you already hold, then score the output against your own cases before anything reaches a customer. Pilot first; production only if the numbers hold.

Atlas Properties, a concept build: a listings dashboard showing 1,240 live listings, filtered results for Business Bay, Dubai Marina and JLT with AED asking prices, a map panel, and an English–Arabic language toggle.

Concept build — Atlas Properties, engineered in-house

1

Task piloted at a time

Agreed

Passing score set before the build

100%

Answers carry a citation

0

Training on your data

What we build

Eight pieces of the job, quoted separately

Eight places a model earns its cost, and where it does not.

How we build it

Three decisions we make early

Three rules that keep a pilot from becoming a demo.

One task, chosen carefully

We rank candidate tasks by two things: what the time is worth, and whether the data behind it is actually readable today. The winner is usually smaller and duller than the one that came up in the meeting, and it is the one that returns money. Everything else waits for the second pilot.

The evaluation set comes first

Before any prompt is written we build a set of your real cases with the answers you would accept, and agree the passing score. That number decides whether this ships. Without it, every review becomes an argument about whether an answer felt right, which is how bad systems reach production.

A person stays in the loop

Where a mistake costs money or reputation, the model drafts and a named person approves. That is a workflow decision, not a limitation — it keeps the speed gain, removes the tail risk, and means you can explain to a regulator or a customer exactly who made the call.

What you get

What a pilot leaves behind, even if you stop there.

  • Search across your own documents, with citations
  • An evaluation set built from your real cases
  • A scoring threshold agreed before the build
  • A human review step wherever mistakes cost money
  • Cost and latency budgets per request
  • Prompts and retrieval logic, handed over as code
  • Data-handling terms you can show your legal team
  • A written recommendation to ship or to stop
  • Infrastructure in your own cloud account
  • Thirty days of warranty after production launch

Stack

What this work actually runs on

What sits between your documents and a model.

Storage, retrieval and models wired to your data.

10 tools

  • PostgreSQL
  • MySQL
  • Redis
  • pgvector
  • Prisma
  • Drizzle
  • OpenAI
  • Anthropic
  • LangChain
  • Metabase

A drawing of an assistant answering: a question in a prompt field, then four lines of a reply arriving left to right, then the sources it used.

Names, not logos. Everything listed is something we build with or integrate, not a category page we copied.

Where it lands

Six sectors this work is bought in

Where the reading-then-doing work is heaviest.

  • Financial services

    Onboarding flows, reporting and audit-ready records.

  • Healthcare

    Clinic systems, patient portals and careful data handling.

  • Government

    Service portals, bilingual forms and accessibility work.

  • Professional services

    Practice management, billing and client portals.

  • Energy

    Field reporting, asset registers and offline-first tools.

  • Education

    Enrolment portals, course platforms and parent apps.

All ten sectors we build for

Process

Five stages, and what you hold at each

Five stages, and a stop point after the pilot.

  1. 01

    Scope

    One task chosen, ranked by value and data readiness.

    Task and threshold

  2. 02

    Prototype

    Retrieval over your own documents, on real cases.

    Working pilot

  3. 03

    Build

    Review steps, permissions and budgets around the model.

    Evaluation report

  4. 04

    Launch

    A limited rollout to one team, watched daily.

    Rollout plan

  5. 05

    Run

    Re-scored monthly, because models and documents both drift.

    Monthly score

How we quote it

We quote the pilot on its own, and production separately — only if the scores hold.

  • The pilot is quoted separately, and you may stop there
  • Prompts, retrieval code and evaluations are yours
  • Defects found in the first thirty days are fixed free

Quoted as ai implementation: data pipeline, retrieval, evaluation, human review loop. We publish no bands — pick the pieces you need and one itemised quotation comes back within 48 hours, free either way.

Get a quote

Questions

Before you send the brief

Where does this actually pay off?

Where someone reads something and then does something predictable: searching an archive, summarising tenders, triaging support mail, extracting fields from invoices, drafting a reply a person still sends. The gain is removing a lookup, not replacing judgement. If the task needs a decision your business would defend in front of a regulator, the model drafts and a named person approves — that is a workflow design, not a limitation.

Will our data train someone else's model?

Not on the endpoints we use. The enterprise and API tiers from the major providers exclude inputs from training by default, and we put that in writing in the scope rather than asking you to trust a blog post. Where the data is sensitive enough that the question keeps coming back, we run through a cloud region you nominate — Azure OpenAI or AWS Bedrock — so the traffic never leaves your tenancy.

How do you know it is good enough?

We build an evaluation set from your real cases at the start, agree the passing threshold with you before development begins, and score every change against it. That number is what the pilot is for. If it fails, you stop at pilot cost and we say so plainly — that outcome is a success for you, because the alternative is finding out after a production rollout.

Does it work in Arabic?

Modern Standard Arabic is handled well by current models; dialect is less reliable, and mixed Arabic-English documents are their own problem. So we test against your own Arabic documents during the pilot rather than assuming, and we index mixed archives together so a search returns both languages. If the scores are weak on your material, that is part of the pilot's finding, not a surprise later.

What does it cost to run each month?

Model calls are usually the smaller line; hosting, the vector store and the humans reviewing output are the larger ones. We set a cost budget per request during scoping and monitor it after launch, because a prompt change can double spend without changing behaviour. You get the arithmetic before you commit, and an alert when a month is trending past what we quoted.

Related

Where this shows up

Two builds where search and data quality decided everything.

  • Property listing platform

    Atlas Properties

    A bilingual property portal: 1,240 listings, faceted search and an agent-facing CMS.

    LCP on a filtered results page
    0.9s
    Lighthouse performance
    100
  • Clinic booking system

    Nadra Clinics

    A four-branch booking system with held slots, WhatsApp confirmations and a front-desk console.

    Lighthouse accessibility
    100
    LCP on the slot picker
    1.1s

Bought alongside

What has to exist before a model is useful.

  • Custom software

    CRM and internal systems shaped around how you already work.

    Explore
  • IT & networking

    Office networks, cloud infrastructure, and support that answers.

    Explore
  • Web development

    Marketing sites and web apps in Next.js, ready for Arabic.

    Explore

Pilot one task first.

Pick the task that eats the most reading time. We will score it against your real cases before you commit to production.