AI for UAE Government & Public Sector

How an air-gapped system stays current (without ever going online)

The usual objection to running offline is that the system goes stale. It does not.

Updates ship the way classified material already moves. A new model build, or a refreshed index, is written to AES-256 encrypted media, hand-carried by cleared staff, and signature-verified before anything installs. Nothing on that media runs until the signature checks out. For a critical CVE, there is an out-of-band path you can choose to use, on your own schedule and your own terms. Re-indexing runs on a set cadence, and supersession is built in, so an updated circular replaces the old one cleanly rather than sitting next to it.

The whole lifecycle stays on your side of the gap. Weights and index live in your custody from day one, and you control when each update lands. The air gap is never bridged to keep the system fresh.

Where government actually sits in UAE data law and assurance

For a public entity, the document that governs you is the UAE Information Assurance Standard, maintained by the Cybersecurity Council and originally issued under NESA, since renamed the SIA. It is about how classified and sensitive information is handled, not just personal data, and a private AI assistant is built to fit it from the start.

The PDPL is a common point of confusion. Under Article 2, government data and the authorities that process it fall outside the law. The exemption is real, but it is not unconditional. The moment you handle non-government personal data, run a public-facing service, or share data with a DIFC or ADGM partner, the personal-data rules can apply again. A deployment is scoped around both the IA Standard and the relevant Management and Technical control families, so the line between the two regimes is drawn before anything goes live.

The trap most procurement teams miss is cross-border transfer. The instant a query or a document touches a foreign-hosted model, your data has left the country, and the assurance picture changes completely. An air-gapped system removes that exposure outright. For this audience, that single property is the strongest argument there is.

What you actually rack: models, hardware, and footprint

Here is the infrastructure, in plain terms, so your IT team can size it before you commit to anything.

One GPU server runs the whole thing. The model is open-weight, somewhere in the 7B to 70B range depending on the work, served with vLLM. A 70B model fits on a single 80-96GB GPU: roughly 38-40GB at Q4, around 70GB at FP8. The Q4_K_M quantisation costs about 3-5% on benchmarks, which is negligible for retrieval-grounded work where the answer comes from your own documents anyway. The result is a self-contained enclave with no external dependencies.

None of this is a guess made in advance. The bill of materials is scoped with your IT team against your actual workload before any hardware is ordered.

How operators trust the answer: citations, confidence, and refusal

An answer you cannot check is an answer you cannot use. So the system is built to be checked.

Every response is grounded in passages pulled from your corpus, and each one carries a clickable citation down to the paragraph. A confidence threshold gates anything sensitive: below the line, the system holds the answer back rather than presenting a weak one. Ask it something the documents do not cover, and it says so plainly instead of filling the gap. The scope is deliberate and stated up front. The assistant retrieves and cites to support a human decision. It does not decide on its own, and it never makes a classification determination.

Working in Arabic and English over your real corpus

Most government corpora are heavily Arabic, often mixed with English in the same file. The system indexes and retrieves across both, answers in whichever language you ask, and cites in the original language of the source. A query that switches between Arabic and English mid-sentence is handled directly, with no lossy translation step in the middle.

This runs on Arabic-capable multilingual embeddings, which in practice score better on Arabic retrieval than models built only for Arabic. On the generation side, the options now include Jais 2 (70B, released December 2025) and Falcon-Arabic from TII. The honest caveat: most open models still carry an English bias, so Arabic is something handled with real care, not a problem anyone should claim is perfectly solved.

How a public entity engages us

This is not an RFP-and-wait process. It is a scoped pilot, run inside your own environment, alongside your security and IT teams.

Deployment is on hardware you own, with the weights and index in your custody from day one. You get the architecture documented, the relevant control families mapped, and the operating procedures written down. Before the engagement ends, your team is trained to run the system without us. The reference point for what this looks like is the prior air-gapped deployment: roughly 8 weeks to a working system, and around 40% less time spent finding the right clause once it was live.