Skip to content

Software

AI development for enterprises and governments

We design and build AI systems that work on your own data and inside your own rules: assistants that answer from internal documents, agents that act with human approval, and the evaluation that shows whether they are right. AI is how we work as well as what we build, so we know where it holds up and where it does not.

Book a callSee how it runs
A brushed gunmetal cube with frosted glass centre tiles, lit from a core that shows through its seams

What it is

AI assistants, retrieval systems and agents with access controls, evaluation and human approval built in.

A demo that impresses in a meeting and a system that an organisation can rely on are far apart. The gap is made of access control, evaluation, failure handling and data governance, and that is where we spend most of our effort. A model call is a small part of the work.

We begin by identifying a task where an AI system has a measurable job: answering staff questions from policy documents, classifying incoming requests, drafting responses for review, extracting fields from forms. We define what a correct answer looks like, build an evaluation set from your real material, and only then build the system against it. Quality is a number you can track rather than an impression. That number is also what lets your compliance and risk teams approve the system with confidence.

Where the data cannot leave a jurisdiction or an environment, we design for that: models hosted in-country or on-premises where the requirement demands it, and hosted APIs where it does not. Arabic and English are treated as first-class from the start. People stay in the loop on anything consequential, and every answer can be traced back to its source. When the model or the data changes, the evaluation is rerun and you see the effect before it reaches users. We keep the evaluation in your repository, so your own team can extend it as new use cases appear.

Capabilities

What the work covers.

  • Retrieval over internal documents

    Assistants that answer from your policies, contracts and knowledge bases, respect each user's access rights, and cite the passage they used.

  • Evaluation harnesses

    Test sets built from real questions and documents, with automated scoring, so changes to prompts or models are measured before release.

  • Agents with human approval steps

    Agents that read, draft and prepare actions across your systems, and pause for a named person to approve anything with consequences.

  • Arabic language handling

    Retrieval, generation and evaluation tested on Modern Standard Arabic and the dialects your users write, including mixed Arabic and English text.

  • In-country and on-premises hosting

    Architectures that keep data and inference inside a required jurisdiction or private environment, with open-weight models where a hosted API is not permitted.

  • Document extraction and classification

    Structured data pulled from scanned forms, invoices and correspondence, with confidence scores and a review queue for uncertain cases.

  • Guardrails and audit logging

    Input and output checks, prompt-injection defences, and a full record of what was asked, retrieved and answered.

  • Model selection and cost control

    An honest comparison of models against your evaluation set, with caching and routing so running costs stay predictable.

At handover

What you keep.

Typical stack

Claude API, OpenAI API, Python, FastAPI, TypeScript, Next.js, PostgreSQL, Drizzle, n8n, Bun, Cloudflare, Railway.

Named plainly. We pick per project and explain why.

  • The deployed AI system and its source code
  • An evaluation set and scoring pipeline you can rerun
  • Documentation of data flows, models used and where data is stored
  • Access-control and audit-log design
  • Prompt and configuration files under version control
  • An operations guide covering monitoring, updates and model changes
  • Training for the staff who will use and supervise it

How it runs

Four steps, every time.

The solved cube on a black stone plinth, its front face showing the BTI logo in brushed metal and frosted glass

  1. Discover

    We map the product, the funnel or the process and agree where the first useful change is.

  2. Design

    Scope, architecture and rollout on paper. You see what ships, in what order, before anything is built.

  3. Build

    Working increments, reviewed with you as they land rather than revealed at the end.

  4. Operate

    We hand over code, documentation and access, train your team, and stay on until running it is routine.

Questions

Straight answers.

How do we choose the best AI development agency?

Look for an agency that talks about evaluation before it talks about models. Ask how it will measure accuracy on your data, how access rights carry through to answers, where your data is processed and stored, and how a person stays in the loop. Prefer a team that has shipped production systems rather than prototypes, can show how it handles failure cases, and will tell you when AI is the wrong tool for the task.

Can the AI run without sending our data outside the country?

Often, yes. Depending on the requirement we can host open-weight models in-country or on your own infrastructure, or use a hosted provider in a permitted region. We map the data flows first and design to the strictest rule that applies.

How do you stop it from making things up?

By grounding answers in retrieved sources, requiring citations, testing against an evaluation set, and having the system decline when it lacks support. For high-stakes outputs, a person reviews before anything is sent or acted on.

Does it handle Arabic?

Yes, and we test for it specifically. Evaluation sets include Arabic and mixed-language questions, and the interface is built right-to-left. Model quality varies by dialect and domain, so we measure it on your material rather than assume.

What does an AI project cost?

It is scoped before kickoff. The main drivers are the number and quality of source documents, the integrations involved, the hosting requirements and the level of evaluation needed. Running costs, such as model usage, are estimated separately and shown to you up front.

Will our staff be replaced by this?

The systems we build take on repetitive reading, searching and drafting so that people spend their time on judgement. We design them with the staff who will use them, and train those staff to supervise the system.

Other services