Pattern 01
Knowledge assistant over internal documents
A permission-aware interface over policies, contracts and manuals. Users only see answers drawn from documents they may read.
02 · Custom AI Applications
Software where the intelligence is part of the product
Applications built around your workflow, with AI doing the part of the job it is genuinely better at, and an interface people can use, trust and correct.
Custom software is worth building when the workflow is genuinely yours. If an existing product already does 95 percent of the job, we will say so, and build only the integration that closes the gap.
Illustrative application · open a document
The screen is where trust is built or lost.
It shows what the model produced, where it came from, and lets a person change it. Try it: nothing is recorded until you confirm.
Accounts payable · illustrative application
Supplier invoice
PDF attachment, 2 pages
- Supplier Read
- Invoice reference Read
- Purchase order Read
- Due date Read
- Matches the purchase order
- Within your approval limit
Model
Read the document
Your systems
Matched your records
Application logic
Checked your rules
Interface
Showed it to a person
Written
Waits for a person
Every edit a user makes is captured. That record becomes the evaluation set, which is why the review screen matters more than the prompt.
When the system should act without waiting for a person, the question becomes what it may do alone. That is an agent, and the two compose well.
01What it is
What this is, concretely
A custom AI application is ordinary production software: authentication, data model, permissions, interface, deployment. Inside it sits model-driven capability at the steps that earn it. Drafting, extracting, ranking, summarising, matching, classifying: work a person would otherwise do slowly and inconsistently.
What separates this from a chat window bolted onto an existing product is the interface. When a model contributes to a decision, the screen has to show what it produced, where that came from, and how a person changes it. That is an interface design problem at least as much as a model problem.
02Anatomy
The model is one layer. The rest is software.
Access, rules, integrations and an audit trail are what make a model's output something a business can use.
- InterfaceScreens people use to review, edit and confirm
- AccessSign-in, roles and per-record permissions
- Application logicValidation, routing and your business rules
- Model boundaryOne controlled place the model is called
- IntegrationsYour records, documents and existing systems
- OperationsAudit trail, evaluation and deployment
03Where it fits
When bespoke software is the right answer.
The workflow lives in spreadsheets
A process that matters, held together by files, inboxes and habit. It works until the person who understands it is away, and it never produces data anyone can analyse.
Off-the-shelf fits at 70 percent
The remaining 30 percent is the part specific to how you operate, and it is the part being handled manually, in parallel, forever.
The AI feature is a demo
Something impressive was built, then stalled on permissions, data access, auditability, or a screen that gives nobody a way to correct it.
04Examples
Systems of this shape
Build patterns and capabilities, not client projects.
Pattern 02
AI-enabled internal platform
The operational tool a team currently improvises, rebuilt properly, with model assistance at the slow and repetitive steps.
Pattern 03
Custom workflow application
A queue, a review screen and an audit trail. The model prepares each item, a person confirms, edits or rejects it, and the edits become the evaluation set.
Pattern 04
Structured extraction pipeline
Documents in, validated records out, with a review screen for the cases the system flags rather than a silent best guess.
05Approach
How we build it.
01
The workflow is the specification
We map the path work takes today, workarounds included, before deciding what software to write. The exceptions are usually where the value is.
02
Deterministic wherever it can be
Validation, calculation, routing and permissions are code. The model handles the parts that genuinely need language or judgement, which keeps behaviour predictable and cost sane.
03
Interfaces that expose uncertainty
Model output arrives as something to confirm, with its source attached, never as settled fact. People trust a system that admits what it is unsure about.
04
Corrections feed back
Every edit a user makes is captured. That record becomes the evaluation set and the honest answer to whether this is actually getting better.
06Data
Where the data lives is an architecture decision.
It is usually the first question a regulated organisation asks, so it is settled before the build, not during it.
01 Where processing happens
Which region the application runs in, which region the model is called in, and whether those are the same. Providers publish their own regional options and they are not all equivalent.
02 Who sees what
Which provider receives which fields, and whether anything identifying has to leave your systems at all. Inside the application, authentication and per-record permissions are ordinary software concerns and the model does not get to bypass them: a user sees answers drawn only from documents they could already open.
03 What is retained, and for how long
Retention of prompts and outputs is a provider setting and a contract term, not a default worth inheriting. It gets decided, written down, and reflected in the architecture.
04 What the record shows afterwards
Which records were touched, by which component, under whose credentials. An audit trail is far easier to design in at the start than to reconstruct once somebody asks for it.
The frameworks that usually apply in our two main markets
- United Arab Emirates
- Federal Decree-Law No. 45 of 2021 on the Protection of Personal Data governs the processing of personal data and sets out requirements for transferring it outside the country. The financial free zones, including the DIFC, run their own data protection regimes alongside it.
- Saudi Arabia
- The Personal Data Protection Law is administered by SDAIA, the Saudi Data and Artificial Intelligence Authority. One point is worth stating precisely, because it is often reported loosely: the law does not require that all personal data stay inside the Kingdom. It permits transfer abroad subject to conditions, which include the level of protection in the receiving country or the use of approved safeguards, limiting a transfer to the personal data actually needed, and not prejudicing national security or the vital interests of the Kingdom. SDAIA publishes separate regulation and guidance covering transfers.
We are engineers rather than your legal advisers, and none of this is legal advice. What we do is build to the position your counsel establishes, put the data-handling decisions into the architecture instead of into a policy document nobody opens, and say so when a design choice would make one of your obligations harder to meet.
07In production
Held to normal software standards
Custom software is worth building when the workflow is genuinely yours. If an existing product already does 95 percent of the job, we will say so, and build only the integration that closes the gap.
It is real software
Typed, tested, version controlled, deployed through a pipeline, with environments and rollback. AI inside does not lower that bar.
Data handling is decided, not assumed
Where data lives, which provider sees it, what is retained and for how long: settled before the build, written down, reflected in the architecture.
The model is replaceable
Provider access sits behind one internal boundary, so changing model or vendor is a configuration change and a re-run of the evaluations, not a rewrite.
It hands over cleanly
Documented architecture, runbooks and access. Operate it yourself, have us operate it, or move it to another team.
08Where and next
Where we do this work
These applications are built mainly for companies in the United Arab Emirates and Saudi Arabia. Where the fit is right we work with clients in the United States and Europe, and where data may live is settled in the same conversation.
Start a projectThe other two lines
- 01AI Agent DevelopmentGive AI the tools, permissions and context to finish real work inside your systems.Explore
- 03Managed AI SupportEngineering operations for AI already in production: evaluation, monitoring, and the change that keeps it correct.Explore
Is the workflow that matters most held together by spreadsheets?
Bring the spreadsheet, the inbox and the person who knows how it really runs. The exceptions in that process are usually the whole specification.