We turn complex challenges into intelligent solutions.
Many companies in the Kingdom still match their documents line by line, by eye. We build systems that read those documents, match them, surface their errors, and leave the final decision to you — with accuracy that holds up to the regulator.
Reading the data is only the beginning. Understanding how it connects is what matters.
Off-the-shelf tools read the text off a document and stop there. The real work starts after reading: does this line on the invoice correspond to that line on the declaration? Does the value match? Is it the same unit, or is the difference cosmetic? And when they don’t match — what kind of error is it, and who decides what happens next?
This is not an OCR problem. It is a matching and decision problem, and that is what we build.
Time goes into checking, not reading
Reading the document takes minutes. Verifying that every field in it matches another document is what consumes your team’s hours, every week.
The error only surfaces after rejection
In regulated operations, errors aren’t caught internally — they’re caught when the transaction is rejected. By then the cost has already landed.
The knowledge lives in people’s heads
One or two people know every edge case and what to do about each. That knowledge isn’t written down anywhere, and it leaves when they do.
The same problem, under different names
This problem repeats across sectors with no apparent connection — but the structure is identical: two or more documents from different sources that must match field by field, where any difference carries a regulatory or financial consequence.
Customs and international trade
Matching the customs declaration against the commercial invoice, HS codes, duty refund files
Accounts payable
Three-way match between the purchase order, the goods receipt and the invoice
Regulatory compliance
Preparing submission files for authorities, and validating them before filing rather than after rejection
Insurance
Matching claims against coverage documents and medical invoices
Contracts and procurement
Comparing contract terms with actual invoices and surfacing deviations
Banking and finance
Verifying account opening documents and letters of credit
If your team opens two documents side by side and compares them by eye — that’s our problem.
A pipeline of separate stages — not a black box
We don’t build a system that swallows a document and emits a result. We build a pipeline where every stage has a visible, stored output. When the system is wrong — and it will be — you know exactly which stage failed, and you fix that stage. A monolithic system is diagnosed by guesswork, and its accuracy can’t be improved because you don’t know where the error came from.
Intake
You upload your documents — PDFs or images, in any shape and from any supplier
Extraction
Every document becomes structured data, with a confidence score per field — not per document
Normalization
Units, currencies, dates and numbers are unified — while the original text is kept exactly as it appeared
Matching
Line items are aligned across documents, then every field is compared against tolerances you set
Classification and decision
On a difference: what kind is it? what is the evidence? what is the suggested route?
Review and output
Your team reviews and decides, and the system emits the file in the approved template
Every value knows where it came from
Every field the system extracts points back to its location in the source document — page and position. Click the number and the document opens exactly there. Nobody is asked to trust a number they can’t trace.
Four decisions that separate a system people rely on from one abandoned after a month
AI reads, code decides
The model is excellent at reading a messy document and understanding its structure. But it is probabilistic — no numeric comparison should pass through it. In our systems: extraction and understanding are AI; comparisons, unit conversion and rule application are deterministic logic that returns the same result every time.
The cost of an error is asymmetric, and the system is tuned for that
A system that flags an error in a clean file costs you minutes of review. A system that lets a file with an error through costs you the whole transaction. The two costs are not equal, so the system must not be tuned as if they were. We tune ours to doubt more than they reassure — we don’t promise a system that never errs, but one that never lets an error through without drawing your attention to it.
The final decision stays with your team
The system classifies, shows the evidence and proposes a route. The human decides. This is not technical caution — it is what makes the system defensible to an auditor or a regulator, and what international standards require for decisions with regulatory effect.
Your rules are data, not buried code
Your business rules will change, and regulations change. So we build them as editable, versioned data — not buried in code. And when a decision made six months ago is challenged, you can produce the rule as it stood at that time.
We train the models and run them — from data generation to production
We don’t stop at calling a ready-made model through an API. When the domain demands accuracy a general model won’t give, we build the right model for your work: we generate and label the training data, tune the model on your documents and your sector’s terminology, measure its accuracy on a held-out test set, then run it, monitor it, and retrain it when the documents change.
Data generation and labelling
We build a training set from your real documents, and synthesize what’s missing with deliberate edge cases.
Tuning and training
Fine-tuning on the structure of your documents and your sector’s terminology, instead of a fixed template per document type.
Measurement
A held-out test set the training never sees, and an accuracy number per field — not per document.
Deployment
Running inside your environment or inside the Kingdom, with confidence thresholds that decide when a field goes to review.
Monitoring and retraining
We watch for accuracy drift when documents change, and retrain on the new cases.
Choosing between a general model and a trained one is a cost and accuracy decision — we make it after measurement, not before.
We don’t ask for a six-month contract before showing anything
We start with the smallest step that proves the idea works on your own documents — not on a slide deck, and not on a sample we picked.
| Stage | Duration | What you get |
|---|---|---|
| 1 — Diagnosis | 30-minute call | A clear read on your problem, and a straight verdict from us: can it be automated at all? |
| 2 — Proof of feasibility | 1–2 weeks | We work through 3 real cases of yours and give you a real accuracy number, plus an analysis of where the system fails and why |
| 3 — Process documentation | 1–2 weeks | All your rules and edge cases documented in writing — value you keep even if we stop here |
| 4 — Build | 6–8 weeks | The full system: extraction, matching, review, outputs |
| 5 — Accuracy tuning | 3–4 weeks | Running in parallel with your manual work until you trust the numbers |
| 6 — Analytics | 2–3 weeks | Dashboards showing where errors repeat and how much you actually save |
We don’t fix a final price before the end of stage 3. Before that we don’t know the real complexity — and anyone who gives you a final price before seeing your edge cases either didn’t understand the problem, or priced it with a margin that covers their ignorance.
Your documents are your business — we treat them that way
| Item | Detail |
|---|---|
| Data location | Your data and documents are hosted inside the Kingdom |
| Full audit trail | Every decision — automated or human — is recorded: what the system proposed, what the human decided, and when |
| Scoped permissions | Who sees what, and who approves what |
| No training on your data | Your documents are never used to train any model, and are not retained by the model provider |
| Full traceability | Every value can be traced back to its source in the original document |
We sign an NDA before you send us the first document.
Candour that saves both sides time
| We don’t promise | Because |
|---|---|
| 100% accuracy from day one | Any vendor who promises that before seeing your documents is selling you an illusion. We give you a real number after two weeks of work on your cases |
| Removing human review | Our goal is to cut ten minutes down to one, not to remove the human from a process that carries regulatory consequence |
| Immediate integration with government systems | Access and permissions are outside our control. So we build the system to work with manual upload first and deliver its full value, then add integration when it becomes available |
| That every problem deserves a system | Sometimes the honest answer is: your volume doesn’t justify this investment yet. We’ll say it in the first call, not three months in |
Frequently asked questions
Our documents are inconsistent — every supplier uses a different format. Is that a problem?
That is the problem itself. Legacy systems need a template per document format, which is endless maintenance. We use models that understand document structure without a predefined template — which is why, in the feasibility proof, we specifically ask for documents from different suppliers: to measure this, not to avoid it.
Some of our documents are poor-quality scans.
Usually workable. But when the system’s confidence in a field drops, it surfaces that field for review instead of guessing silently. Silent failure is more dangerous than declared failure.
What does it cost?
It depends on how many edge cases you have, which isn’t known before stage 3. The diagnostic call is free, and we’ll give you an estimated range in it.
Do we need a technical team on our side?
No. We need one person who knows the manual process well, can answer our questions and review the output weekly. That person is the most important resource in the project — more than anything technical.
What if we stop after the feasibility proof?
You walk away with a real accuracy number on your own documents, and a clear understanding of whether your problem can be automated. That knowledge stays with you.
Do you work with companies outside the Kingdom?
Yes, but our regulatory experience is concentrated in the Saudi market.
The bodies and platforms we are registered with
Programmes, platforms and accreditations registered to us as a company — government bodies, infrastructure and cloud providers.
Commercial registration and compliance documents are shared on request.
Start with a diagnostic call — 30 minutes
Tell us about the manual process eating your team’s time. We’ll tell you honestly whether it can be automated, what that would take, and a preliminary estimate. No sales pitch.