Ministries and public authorities hold handwritten Arabic application forms, permits, court files and decades of scanned archive — on networks that frequently have no outbound route by policy. DocxIntel installs inside that perimeter with the model weights in the bundle and no foreign cloud dependency.
Accuracy is measured field-by-field against a human-adjudicated ground truth, not character-by-character. Handwritten Arabic forms are part of the published benchmark mix, and your own threshold is agreed in writing on your own documents before the Proof of Value begins.
Public services digitised the front end first. The back office still receives handwritten forms at a counter, and the records room still holds the shelves nobody has been able to index.
Handwritten Arabic service requests are read into structured fields — applicant identity, service requested, dates, declarations and signatures — with uncertain entries flagged for a clerk.
Permits, licences and their renewals are extracted and validated against the register, including expiry dates, conditions attached and the issuing authority.
Decades of scanned records — including poor microfilm scans, faded typescript and multi-generation photocopies — are converted into searchable, classified records.
Case files, judgments, pleadings and hearing records are classified, indexed by party and case reference, and made searchable with citations back to the page.
Tender submissions, bid tables, technical compliance matrices and supporting certificates are extracted so evaluation panels compare structured data rather than PDFs.
Documents are classified into your own records categories on ingest, making retention schedules enforceable and disclosure searches answerable with citations.
Backfile conversion is the project every authority has costed and shelved, because metered processing makes the page count the budget. On a fixed licence the page count is a scheduling question.
This runs identically on an air-gapped network. Nothing about the workflow assumes an outbound connection exists.
Inside your own accreditation boundary, on your own documents, with your own team holding the keys throughout.
A single high-volume counter service, or one archive series, is a better first scope than a department. We ask for the intake your clerks find hardest — the handwritten forms, the faded archive, the poor microfilm — because that determines the real number.
Network zone, isolation requirements, logging destinations and operator roles are agreed with your information assurance team before installation. Deployment evidence and data-flow documentation are provided for the accreditation file.
Containers and open-weight models are installed from media, on hardware you own, with no repository access required. Your team holds root, the encryption keys and the change window from the first day.
Your form layouts, field schemas, records-classification categories and retention schedules are configured directly. No template library is used, and no sample document leaves the environment it was loaded into.
Your clerks and records officers adjudicate field-by-field, and the report shows per-field accuracy, confidence distribution and the review load you would actually staff. Miss the agreed threshold and you keep the report and owe nothing further.
Adding another service, another archive series or millions more archived pages is a matter of GPU capacity and scheduling. The licence does not move because the page count did.
An indicative schema. The final field list is agreed against your own forms, records categories and retention rules during the Proof of Value.
Deployment against a specific accreditation boundary is confirmed in writing during the Proof of Value against your actual network and document mix, not against a generic architecture.
Cloud IDP platforms are competent software sold through hyperscaler marketplaces. For a sovereign authority the question is not capability but jurisdiction, and for an archive project it is arithmetic.
Where documents are read
Works with no outbound route
Foreign cloud dependency
Handwritten Arabic intake
Converting 2M archived pages
Reprocessing after a model improvement
Chain of custody and audit logging
Who holds the model weights
Cost figures are our own arithmetic on publicly published list rates as of 2026, not vendor quotations, and exclude scanning and quality-assurance effort in all three columns. Verify against current vendor documentation before making a decision.
The questions that decide whether document automation can be accredited at all, answered without hedging.
Yes, and that is the primary deployment model rather than a hardened variant. The models ship as open weights inside the deployment bundle and load from local storage, so an air-gapped install performs identically to a connected one. There is no licence check-in, no telemetry beacon and no fallback call when confidence is low — there is nothing outbound to firewall.
Thirty to forty-five days, fixed fee, deployed inside your own accreditation boundary on your own documents. The accuracy threshold and the conversion price are both agreed in writing before we begin.
