العربية Start a conversation

Egypt's data protection law: compliance due from 1 November 2026. What changes, and the checklist

Orchestra · an Arabic LLM that runs on-premises

Our Arabic model. Your servers. One fixed price.

Orchestra is an Arabic language model we built ourselves. It runs inside your network with no outside service it depends on, for one annual licence instead of a bill per token. And nobody abroad can switch it off.

We built it from scratch because the Arabic models we tested read formal Arabic well and struggled with the Arabic customers actually write. We're measuring how much better Orchestra does, and we'll publish the results.

Talk to one of our engineers About Orchestra

Headquartered in Cairo · partners in Dubai and Riyadh

Diagram: Orchestra running on a server inside the organisation's network with no outside service; an assistant answering in Gulf Arabic with two cited sources; a fixed annual licence, zero tokens billed, no data leaving the network.
Products

Orchestra, on your servers.

An Arabic language model our engineers install inside your network, with search over your documents, a safety layer and logs.

  • 14 billion parameters, built from scratch at NoorAI
  • For formal Arabic and the Arabic customers actually write
  • On your own GPU server, air-gapped when required
  • One fixed annual licence, no token bill

Orchestra

An Arabic language model we built ourselves. Our engineers install it on your servers with search over your documents, a safety layer and logs, and all of it stays inside your network.

About Orchestra

By industry

Banking and financial services

Customer service in customers’ dialects, complaint triage and policy search, without customer data leaving the bank.

AI for banking

Government

Citizen services, official correspondence and archive search, inside the network and air-gapped when required.

AI for government

Healthcare

Patient appointments, documentation drafts and insurance requests, without health data leaving the hospital.

AI for healthcare

Telecom

Customer care in every dialect, contact-centre agent assist and call summaries, inside your network.

AI for telecom

We also run LeadX, a hosted sales assistant that builds lead lists, available by request. About LeadX

Security

Five layers, with a security control at each one.

This is what your security team usually asks about, layer by layer. Every step is logged and can be reviewed, and sensitive actions pass by a person before they run.

  1. IngressWhatsApp, web, email and voice, behind enterprise identity (SAML, OIDC)
  2. ModelOrchestra served air-gapped on your hardware, with a safety classifier and auditable policy rules
  3. RetrievalQdrant hybrid search that respects access rights, with sources cited in every answer
  4. ActionTool use through the Model Context Protocol, with human review for sensitive actions
  5. OperationsKubernetes and Terraform on your infrastructure, with versioning, regression tests, full tracing and immutable audit logs
audit trace · whatsapp-agent · example
Example audit trace of one conversation through the five layers
TimeEventStatus
09:14:02.118ingress.whatsapp identity verified (OIDC)ok
09:14:02.164safety.classify policy checkpass
09:14:02.391retrieval.hybrid 4 chunks · ACL filteredok
09:14:02.902mcp.tool crm.lookup_customerok
09:14:03.410orchestra.generate on-prem · 2 citationsok
09:14:03.415review.route refund request → humanescalated
09:14:03.420audit.write immutablesealed
VersioningRegression testingHuman escalationFull tracingImmutable audit logs
Deployment

Deploy where your data lives.

We deliver the full stack inside your environment, evaluate it before launch, and run it with you afterwards. Same stack and same security controls in every option. What is an on-premises LLM?

Your data center

On-premises

  • Runs on your servers and network
  • Integrated with your identity provider
  • Connected to internal systems through MCP
Plan an on-prem rollout
Maximum isolation

Air-gapped

  • No network path out
  • Weights, indexes, logs and updates stay inside
  • Built for government, regulators and banks
Start a conversation
In-country

Dedicated private GPUs

  • Single-tenant inference, no shared capacity
  • In an in-country data center
  • Same security controls and audit trail
Talk to an engineer
  1. Phase 1Discover

    Use-case scoping, data and infrastructure audit, compliance mapping.

  2. Phase 2Deploy

    Three environments: staging, parallel run, production. Evaluation suite signed off before launch.

  3. Phase 3Operate

    Monitoring and evaluation, agents expanded as they prove reliable, team training.

Why our own model

We couldn't find a light Arabic model that runs on-premises, so we built one.

The models we tested read formal Arabic well. But customers don't write in formal Arabic. They write like this:

  • كيف حالك؟Modern Standard
  • شلونك؟Gulf
  • إزيك؟Egyptian
  • كيفك؟Levantine

We built Orchestra from scratch for Arabic in all its forms, to run on your own GPU server inside your network. We are now measuring how it does on dialects against the available models, and we'll publish the results when the measurement is complete. About Orchestra

Blog

From the blog.

Data rules in Egypt and the Gulf, and what to check before you deploy Arabic AI.

Security

An AI assistant should only know what you know

An assistant that searches every company document under one account can show people what they should never see. How we apply each document’s permissions inside the search, and why an assistant acts with the permissions of the person it serves.

All posts

Contact

Talk to us.

Tell us what the system should do and where the data must stay, and one of our engineers will reply.

contact@nooraisolutions.com

Or book a 30-minute intro call