NoorAI Solutions — Arabic-first enterprise AI
Everyone else integrates.
We build.
Proprietary Arabic-native language models, an agentic automation platform, and sovereign on-premises deployments — engineered from the ground up for enterprises in the Gulf and beyond.
Who we are
An applied AI lab with production deployments — not a wrapper company.
NoorAI owns the layers of technology most AI companies rent. Our language models are trained by us. Our retrieval, safety, and orchestration stack is assembled and hardened by us. Every layer is self-hosted and independently swappable.
That independence is the point: when your AI has to live inside your data center, comply with your regulator, and speak your customers' dialect — you need a builder, not a broker of someone else's APIs.
What we build
Optimize — Arabic-first LLM
A 14-billion-parameter multilingual language model built from scratch, tuned for Modern Standard Arabic and Gulf dialects, and served fully air-gapped on your infrastructure.
The model PlatformAgentic Platform
A low-code platform for building and operating enterprise AI agents — ingress to action — with orchestration, observability, and enterprise identity built in.
The platform ServicesEnterprise AI Services
Forward-deployed engineering for regulated industries: sovereign deployments, retrieval systems over your documents, and agents wired into your real workflows.
The servicesWhy Arabic-first matters
Most AI treats Arabic as a translation target. We treat it as the native tongue — Noor means light, and it starts at the source.
- DialectsMSA plus Gulf, Egyptian, and Levantine coverage — trained on curated Arabic corpora, not translated English data.
- SovereigntyData never leaves your country. Full stacks deployed inside customer data centers across the GCC, with staging, parallel-run, and production environments.
- ComplianceBuilt for regulated industries — layered safety classification and guardrails, auditable retrieval grounding, and explainable outputs.
Research
An AI lab, not a reseller.
We train neural codecs from scratch, publish on speech representation learning, and run an Arabic-first pretraining program most companies our size wouldn't attempt. That depth is why the products work.
Explore our researchWhere we operate
Dubai. Riyadh. Cairo.
Engineering and delivery across the UAE, Saudi Arabia, and Egypt — part of the Lavaloon group, close to the enterprises we serve and the regulators they answer to.
Next step
See what sovereign Arabic AI looks like inside your walls.
A first conversation takes thirty minutes. A pilot takes weeks, not quarters.
Start a conversationTechnology
Owned at every layer.
Two proprietary systems anchor everything we do: the Optimize family of Arabic-first language models, and the agentic platform that puts them to work. Each has its own story — start where your problem starts.
Optimize — the Arabic-first model family
A 14-billion-parameter flagship trained from scratch, the serving and retrieval stack around it, and the safety layer regulators can inspect. How we train it, how we ground it, how it runs inside your walls.
Explore Optimize The platformThe agentic platform
Five layers from a message arriving to an action taken — low-code agent building, open tool protocols, enterprise identity, and observability across every conversation. Proven at 33K+ conversations a month.
Explore the platformEngineering choices
Boring where it should be. Novel where it counts.
- LanguagesTypeScript for platform services, Python for ML — each where it's strongest.
- DataPostgreSQL, Redis, and MongoDB for platform state; Qdrant for vectors.
- ProtocolsModel Context Protocol for tool integration — open standard, no lock-in.
- IdentitySAML and OIDC so agents inherit your existing enterprise access control.
- ResearchActive work in speech and language modeling, with results targeted at peer-reviewed venues — explore our research.
The short version
We didn't assemble a stack from other people's parts. We built the parts.
Technology — Optimize
The Arabic-first model family.
Optimize is a proprietary family of language models with Arabic as the anchor language, not an afterthought. The flagship, Optimize 1.3, carries 14 billion parameters — trained, aligned, and served entirely by us.
Design
Arabic decides the architecture.
Most multilingual models treat Arabic as one language among a hundred, tokenized through a vocabulary designed for English. Optimize starts from the other end: a tokenizer built for Arabic morphology, a training corpus curated for Modern Standard Arabic and Gulf, Egyptian, and Levantine dialects, and evaluation suites that measure what public benchmarks miss.
The result is a model that spends its capacity on your language instead of wasting it re-assembling fragmented tokens — better quality per parameter, and meaningfully cheaper inference on Arabic workloads.
Specification
- FlagshipOptimize 1.3 — 14B parameters, multilingual with Arabic as the anchor language.
- LanguagesModern Standard Arabic, Gulf, Egyptian, and Levantine dialects, plus English for mixed enterprise environments.
- ContextLong-context configuration for enterprise documents — contracts, filings, and multi-hundred-page regulatory submissions.
- ServingvLLM-based high-throughput inference on customer GPUs, tuned for Arabic tokenization efficiency.
- DeploymentAir-gapped on-premises, private cloud, or dedicated VPC — always in-country, never a shared endpoint.
- EnvironmentsThree per engagement: staging, parallel run, and production.
Training
We train, not just serve.
Owning a model means owning the machinery that improves it. This is the loop that runs behind every Optimize release.
Corpus & data pipelines
Proprietary curation of Arabic text at scale — deduplication, quality filtering, dialect balancing, and domain enrichment from legal, financial, and government registers.
Distributed pretraining
Multi-GPU training with DeepSpeed and FSDP, with every run tracked in MLflow from data mix to final checkpoint — reproducible, auditable, comparable.
Instruction tuning & alignment
Supervised fine-tuning and preference alignment on Arabic instruction data, tuned for the registers enterprises actually use — formal correspondence, regulatory language, customer dialogue.
Domain evaluation
Per-deployment evaluation harnesses. Arabic legal language behaves nothing like Arabic customer support, so pass criteria are defined against your documents, not a public leaderboard.
Grounding
Every answer shows its sources.
Optimize ships with a proprietary retrieval pipeline built on Qdrant: hybrid dense and keyword search over your document estate, with citation-grounded generation so every claim traces back to a source your auditors can open.
- IngestionArabic-capable OCR and structuring pipelines for scanned archives, forms, and decades of inconsistent formatting.
- RetrievalHybrid dense + keyword retrieval with access-control awareness — the model only sees what the asking user is allowed to see.
- Knowledge graphsNeo4j-backed graph retrieval available where entity relationships carry the value; deliberately deferred where vector retrieval alone earns the result.
Safety
Guardrails your regulator can read.
Safety is layered, not bolted on: a dedicated safety classifier screens inputs and outputs, and programmable policy rails encode your compliance requirements — not a generic Western policy set translated after the fact.
Every policy decision is logged and explainable. When a regulator asks why the system refused, redacted, or escalated, the answer is a rule they can read — not a shrug at a black box.
Next step
See Optimize evaluated on your documents, in your data center.
Technology — Agentic platform
From message to action.
The platform is where Optimize goes to work: a low-code system for building, governing, and operating enterprise AI agents — proven in production at more than 33,000 conversations a month.
Architecture
Five layers, every hop observable.
A request enters at the top and leaves as a completed action at the bottom — with every credential governed and every step traceable.
Ingress
Channels where work arrives — WhatsApp, web, email, voice, and internal systems — normalized into a single event stream with per-channel rate control and identity resolution.
Intelligence
Optimize models plus retrieval and safety: understanding the request, grounding it in your knowledge, deciding what should happen — and knowing when to hand off to a human.
Action
Tool use over the Model Context Protocol into your ERP, CRM, ticketing, and databases — agents that do things, not just say things, with every tool call scoped and logged.
Platform
The low-code builder, agent versioning, evaluation suites, human-in-the-loop review queues, and observability across every conversation and tool call.
Infrastructure abstraction
Kubernetes and Terraform underneath; SAML/OIDC identity on top. The same platform runs on your cloud, our cloud, or bare metal behind your firewall.
Building agents
Low-code for the team. Full-code for the edge cases.
Business teams compose agents from building blocks — knowledge sources, tools, escalation rules, tone. Engineering drops to code when a workflow needs it. Both paths land in the same versioned, evaluated, auditable agent definition.
- VersioningEvery agent change is a version — diffable, reviewable, and instantly revertible in production.
- EvaluationRegression suites run against every change: real past conversations replayed before a new version ships.
- Human-in-the-loopConfidence-based escalation into review queues, so the agent earns autonomy case type by case type.
Operations & trust
Governed like the enterprise system it is.
- IdentitySAML and OIDC — agents inherit your existing access control; permissions follow the asking user, not the agent.
- ObservabilityFull traces of every conversation, retrieval, and tool call — searchable, exportable, and retained on your terms.
- AuditImmutable logs of agent decisions and policy outcomes, built for the compliance review, not retrofitted for it.
- StatePostgreSQL, Redis, MongoDB for platform state; Qdrant for vectors — all self-hosted alongside the model.
- DeploymentKubernetes + Terraform — reproducible installs on your cloud, our cloud, or air-gapped bare metal.
Next step
Put an agent on your busiest channel and measure it honestly.
Research
Research feeds product.
Product funds research.
We run an applied research program in Arabic language and speech modeling. The findings ship into Optimize and the platform first, and into peer-reviewed venues second.
Selected work
Codebook Diversity Loss for neural audio codecs
A training objective addressing four failure modes we identified in discrete speech representation learning: semantic codebook collapse, binary acoustic code saturation, training–inference mismatch, and embedding sensitivity. Developed while training a 149M-parameter codec encoder from scratch, and validated against production Arabic speech workloads.
End-of-turn detection for dialectal Arabic conversation
Knowing when a speaker has finished is the difference between an agent that feels attentive and one that interrupts. Our work covers turn-boundary prediction across Gulf, Egyptian, and Levantine speech patterns, where pause structure and discourse markers diverge sharply from English assumptions.
Arabic-first pretraining for the Optimize model family
Corpus curation, tokenizer design, and data-mix strategy for training multilingual models where Arabic is the anchor language rather than an afterthought — including evaluation suites for MSA and dialect performance that public benchmarks don't capture.
Grounded generation for regulated Arabic domains
Retrieval-grounding and citation strategies that hold up under audit: how to make a model's every claim traceable to a source document when the documents are Arabic legal and financial text, scanned at varying quality, spanning decades of formatting conventions.
Research areas
Four fronts, one language.
Everything we study serves the same thesis: Arabic deserves models built for it, not adapted to it.
Arabic LLM pretraining & alignment
Data curation, tokenization, instruction tuning, and preference alignment for MSA and dialects — the full pipeline behind the Optimize family.
Neural codecs & speech representation
Discrete audio representations trained from scratch, and the loss-function research that keeps their codebooks alive and expressive.
Conversational dynamics
Turn-taking, end-of-turn detection, and latency budgets — the unglamorous work that makes an agent feel like a colleague instead of a form.
Grounding, safety & evaluation
Citation-faithful generation, Arabic-tuned safety classification, and domain evaluation harnesses regulators can actually inspect.
How we publish
Specific or silent.
We publish when we have something measured to say, and keep quiet when we don't. Results target peer-reviewed venues; techniques with direct competitive value stay proprietary and show up as product performance instead.
If you're working on Arabic NLP or speech and want to collaborate — or you want your problem to become our next research question — talk to us.
Services
Forward-deployed, not fly-in-fly-out.
Our engineers work inside your environment and your constraints. Consultants leave a deck; we leave a running system your team can operate.
Sovereign AI deployment
End-to-end delivery of the full NoorAI stack — model, retrieval, guardrails, and platform — inside your data center. Air-gapped where required, with staging and parallel-run environments so go-live is a formality, not a leap.
- On-premises
- Air-gapped
- Staging → parallel run → production
- GPU sizing & procurement guidance
Enterprise knowledge & RAG systems
Retrieval systems over your real documents — Arabic and English, scanned and born-digital. Verification workflows, grounded question answering, and search that respects your access controls and cites its sources.
- Document verification
- Arabic OCR pipelines
- Grounded Q&A
- Hybrid retrieval
Agentic automation
Agents deployed on the channels your customers already use — WhatsApp, web, voice — and wired into the systems your teams already run. Scoped narrow, measured honestly, expanded once they earn it.
- WhatsApp & web agents
- Workflow automation
- Human-in-the-loop
- Tool integration via MCP
AI advisory & enablement
Architecture reviews, model selection, compliance mapping, and hands-on enablement for your engineers — so the capability we build stays in your organization after we hand over the keys.
- Architecture review
- Regulatory alignment
- Team enablement
- Evaluation design
How an engagement runs
Three phases. Weeks, not quarters.
Every engagement follows the same spine, sized to your scope. The numbering is the schedule.
Discover
We map your use case, data, infrastructure, and regulatory constraints — and define the evaluation criteria the system must pass before anyone calls it done.
Deploy
The stack goes into your environment: staging first, then a parallel run against your live workload with your team reviewing every output.
Operate
Production cut-over with monitoring, monthly support, and a continuous improvement loop — prompts, retrieval, and models sharpened on real usage.
Selected engagements
Named privately.
Described honestly.
Our customers operate in regulated markets, so we don't put their logos on a wall. What we can share is the shape of the work.
Automated fund-document verification for a national regulatory program
A fully air-gapped Arabic LLM stack — Optimize, retrieval, OCR, and guardrails — verifying fund documentation against regulatory requirements inside the customer's own data center, with staging and parallel-run environments proving accuracy against human reviewers before cut-over.
- Air-gapped on-premises
- Arabic document intelligence
- Audit-traceable outputs
Group-wide AI deployment for a Saudi multi-industry conglomerate
A sovereign deployment serving multiple operating companies from shared infrastructure — one model estate, one governance layer, per-entity isolation — designed around a 36-month total-cost-of-ownership model rather than open-ended cloud spend.
- Multi-entity architecture
- Shared sovereign infrastructure
- 36-month TCO model
Dedicated inference infrastructure for a UAE enterprise
Private high-throughput serving of the Optimize family on customer-controlled GPUs — capacity planning, tokenizer-level Arabic efficiency tuning, and an SLA the customer's own platform team can hold us to.
- Private GPU serving
- vLLM at production scale
- In-country data residency
Production WhatsApp agent at 33K+ conversations a month
A self-hosted conversational agent on our platform handling tens of thousands of monthly customer conversations end-to-end — retrieval-grounded answers, human escalation paths, and monthly tuning cycles that sharpen prompts and knowledge against real usage.
- 33K+ conversations / month
- Self-hosted model serving
- Human-in-the-loop escalation
Who we serve
Regulated first.
Financial services, government, healthcare, telecom — organizations where "just call an API" is not an option, where Arabic is the language of the customer and the regulator, and where data residency is law rather than preference.
If your last AI project stalled on compliance, data egress, or Arabic quality, that's the problem we were built for.
Company
Light, engineered.
Noor is the Arabic word for light. We chose it because that's the job: making enterprise knowledge visible, in the language it was written in.
The thesis
The Arab world will not rent its intelligence. The next decade of enterprise AI in the region will be sovereign, Arabic-native, and owned — and someone has to build it.
- The marketHundreds of millions of Arabic speakers, and regulators writing data-residency into law. Gulf enterprises are mandated to modernize and constrained from sending data abroad — a gap no foreign API can fill.
- The failure modeArabic-after doesn't work. Models designed for English and localized later underperform on dialect, on script, and on the legal and financial registers where the money actually moves.
- Our betOwn the full stack. Models, retrieval, safety, and platform built in-house means every layer can live inside a customer's perimeter — and every improvement compounds into an asset we own, not a bill we pay.
Vision
The AI infrastructure layer of the Arab enterprise.
In ten years, when a bank in Riyadh, a ministry in Abu Dhabi, or a conglomerate in Cairo runs intelligence inside its own walls, we intend for it to run on our stack. Not the most talked-about AI company in the region — the one holding the most production workloads.
The path there is unglamorous by design: win regulated deployments others can't serve, let each one deepen the models and the platform, and publish research when we have something measured to say.
Mission, in one line: make Arabic a first-class language of machine intelligence — and make owning that intelligence the default for the region's institutions.
What we believe
Ownership over integration.
The AI market is full of companies reselling the same handful of Western APIs with a new logo. That model breaks the moment a customer says: the data cannot leave this building, and the model must understand our dialect.
So we took the harder road — training our own models, building our own retrieval and safety stack, and owning the platform end to end. It is slower to build and better to own, for us and for every customer who deploys it.
NoorAI Solutions operates as part of the Lavaloon group, combining Lavaloon's enterprise delivery footprint with a dedicated applied-AI engineering team.
Principles
- Arabic-firstNot Arabic-after. Dialect, script, and culture are design inputs from day one — never a localization pass at the end.
- SovereignYour data stays yours. We design for the strictest residency requirement in the room, then everything easier follows.
- SpecificNo revolutionary anything. We publish numbers, name our stack, and let technical specificity carry the credibility.
- AccountableEvaluation before celebration. Every deployment defines pass criteria up front and runs in parallel against reality before cut-over.
Where we operate
Three cities, one perimeter: yours.
Headquartered in Dubai, delivering across Saudi Arabia, with research and operations in Cairo.
Contact
Start a conversation.
Tell us about your use case, your constraints, and your timeline. We'll come back with a concrete point of view — not a pitch deck.
Offices
- DubaiUnited Arab Emirates — headquarters & engineering
- RiyadhSaudi Arabia — enterprise delivery
- CairoEgypt — R&D & operations
Direct
- Emailhello@noorai.solutions
- GroupA Lavaloon company