Private AI workspace, project management, and API, for the UAE and the GCC
Sovereign AI,
served securely.
Your team works on real, confidential data with the models you approve. Keep everything on local models in the country, or add Claude, GPT, and Gemini with personal data masked before anything leaves. Every request is logged, so you can show a regulator where it went.
The question that decides the sign-off
Data residency is where the files sit. Inference residency is where the GPU runs the prompt. A regulator asks about both.
A "UAE region" on the order form
Content is stored in the country. The prompt can still be processed by a GPU somewhere else, and routing or analytics can run abroad. For the seconds the model is working, the data has left. Ask where inference runs before you sign.
Dalla
In local-only mode, storage, inference, logs, and embeddings stay in the country. Add global models and personal data is masked in the country first, so the model only sees stand-ins. The audit log records the route of every request, so "where did it go" is answered with a report, not a promise.
The workspace
What you need around the models to actually run them
A raw model API is a prompt box. A bank, hospital, or ministry needs control over where data goes, masking, keys, budgets, and a trail. That layer is the product.
You choose what leaves
Run local models only, hosted in the country or on your own GPUs, and nothing crosses the border. Or add Claude, GPT, and Gemini for the hardest work, with personal data masked first.
Masking for global models
Emirates IDs, IBANs, mobile numbers, names, and similar fields are swapped for format-valid stand-ins, in English and Arabic. The model works on the stand-ins, and the answer comes back with the real values restored. On Arabic test cases written after tuning, every value was masked and none leaked.
Private ChatGPT for staff
A chat interface people already know, with a line under each answer that says what was masked and which model ran.
OpenAI-compatible API
Developers keep LangChain, scripts, and internal tools. Change the endpoint and the key. An Anthropic-compatible API is there too.
Routing that holds cost
Simple prompts go to cheaper models. Hard ones go to the strongest model your policy allows. Users can pin a model when it matters.
Keys, budgets, and an audit trail
Each person or team gets a key with a spend cap. Every request is logged with user, model, tier, cost, and route, in a form a PDPL review can read. Admins can switch any model off in one click.
- Request
- req_01j8…k4
- Key
- team-legal · within budget
- Mode
- local + global
- Masked
- 3 fields Emirates ID IBAN name
- Route
- auto → tier 3 · claude
- Sent out
- stand-ins only
- Retention
- record only · no prompt body
- Cost
- AED 0.14
How it works
A question in, a provable answer out
Ask
A person types in chat, or an app calls the API with its own key.
Mask
Before a prompt goes to a global model, personal data is replaced with stand-ins that keep the format. The originals stay in a vault inside the perimeter.
Route
Dalla scores the masked prompt and picks the right model for it.
Run
A local model runs in the country. A global model gets the stand-ins only.
Prove
Real values are restored in the answer. User, model, cost, and route go to the audit log.
Beyond chat
What your teams do in it
The same masking, routing, and audit trail cover everything below, not only the chat box.
Knowledge with RAG
Upload policies, contracts, and reports, including scanned PDFs, Word, Excel, and PowerPoint. Retrieval-augmented generation (RAG) finds the right passages, and answers cite the source. Knowledge can be private, shared with a project, or open to the whole organization.
Assistants
Build an assistant with its own instructions, knowledge, and skills, then share it with people, a project, or everyone.
Prompt library
Shared prompts with folders and fill-in blanks, for one person, a project, or the whole workspace. Type / in the chat to use one.
Automations
Run a prompt on a schedule, from a webhook, or when something happens in the workspace, with a history of every run.
Meetings
Upload a recording and get a transcript and notes with action items, masked like everything else. Action items can go straight onto a board. Staff can also talk to the chat instead of typing.
Files and web search
Create Word, PDF, and Excel files from a chat. Search the web with the query masked before it goes out.
Tools through MCP
Connect MCP (Model Context Protocol) servers once. Each group of tools is on, ask first, or off, and tool inputs are masked like prompts.
Roles, projects, and security analytics
Custom roles, groups, and projects decide who uses which models and data. Probes flag or block ID numbers, secrets, and confidential terms, and admins review them in one place.
Work
Projects, forms, and approvals, inside the same perimeter
Work is project management built into Dalla. Boards, docs, and forms sit in the same instance as your chats, with the same sign-in, roles, and projects. Wherever Dalla runs, in your tenant, your data centre, or our UAE cloud, your project data is there too.
Boards for any team
Table, kanban, calendar, timeline, workload, and dashboard views of the same items, with people, status, dates, files, formulas, and links between boards. A doc can hold a live board, so the brief and the plan stay in step.
Dalla on the team
Assign Dalla a task like anyone else. It works with what you can see, comments what it did, and says so when it can't do something. It doesn't phone or message anyone outside.
Ask about any board
Ask what's stuck and who owns it, have Dalla update items, or describe a board and it drafts one. It proposes, you confirm, and you can take it back.
Approvals that hold
A status change can wait for a named approver, and Dalla's own changes wait the same way. A rejection asks for a reason and lands as a comment.
Forms for the public
Branded forms anyone can fill in without an account, in English or Arabic, with follow-up questions that only show when needed. Each answer lands on a board, and a rule decides who picks it up.
Guests see only their part
Share one board with a client or contractor. Columns you hide never reach their browser, and they can comment only if you allow it.
When Dalla reads a board or a form, it goes through the same masking and audit log as a chat. An Emirates ID on a clinic's intake form reaches a global model as a stand-in, and in local-only mode it doesn't leave the country at all.
Your instance
Configured for your deployment
Every customer gets an instance set up for their rules. We agree these settings with your team before go-live.
Models
Local models only, or local plus global. Bring your own, such as Jais or Falcon on your GPUs, behind an OpenAI-compatible endpoint.
Masking
On or off, which kinds of data it covers, your own list of confidential terms, and an allow list for names that are public.
Retention
How long chats and files are kept, down to zero data retention (ZDR), where no prompt is stored at all.
Single sign-on
Microsoft Entra ID or any SAML provider, with users and groups kept in sync (SCIM).
Audit export
Stream the audit log to your SIEM (security information and event management), such as Microsoft Sentinel or Splunk.
Guardrails
Prompt-injection and jailbreak detection, never-send rules, and model allow lists per group or project.
Deployment
Three ways to run it, one perimeter
Most teams run Dalla in their own cloud tenant. Managed hosting is the fastest start, and on-premise covers networks with no egress. Same product, same log.
Your cloud tenant
Installed inside your own Azure or other cloud subscription, in your country. Your identity, your keys, your bill.
Managed by us
Hosted by us on UAE cloud and set up for your rules. The fastest way to start, with local models from day one and global models with masking if you want them.
On-premise or air-gapped
Runs in your data centre with no egress, on the models you host.
Outside the UAE? Teams in Saudi Arabia, Qatar, Bahrain, Oman, and Kuwait run the same install in their own tenant or data centre, so the perimeter is theirs, in their country. Teams outside the region are welcome too.
Works with what you have
Runs on the cloud and models you already trust
Dalla is the workspace layer. It sits on top of the infrastructure and models your organization has approved, and adds chat, knowledge, masking, keys, budgets, and the audit trail.
Sovereign UAE platforms
Use the in-country models and compute you already contract for, such as Core42 Compass.
Your cloud
Run Dalla in your Azure tenant, including UAE North, next to the Azure OpenAI deployments you have approved. Other clouds and regions in the GCC work the same way.
Your own GPUs
Jais, Falcon, Llama, Qwen, or any model you serve behind an OpenAI-compatible endpoint, such as vLLM.
Global frontier models
Claude, GPT, and Gemini, with masking, when your policy allows them.
Who it is for
Built for teams that cannot send data out
Teams in the UAE and across the GCC that answer to a regulator, and cannot send customer, patient, or citizen data abroad.
Banks and financial services
Contracts, policies, and customer records, with account numbers and IDs masked before any model sees them.
Hospitals and healthcare
Patient notes and clinical documents on local models, under your own access controls. Intake forms land on a board, with Emirates IDs masked before any global model reads them.
Government and public sector
Citizen and classified data inside the country while staff use a private ChatGPT-style workspace, or an air-gapped install.
Legal and professional services
Case files, contracts, and client material without breaking client confidentiality. Share a matter's board with the client, with the columns they shouldn't see held back.
Common questions
Using AI on confidential data, in the UAE and the GCC
What compliance and technology teams ask us most: PDPL, local and global models, masking, your own documents, single sign-on, on-prem, and zero data retention.
Can we use Claude, GPT, or Gemini in the UAE?
Yes, in a way your compliance team can sign off. Dalla masks personal data in the country before a prompt goes to a global model, so Claude, GPT, or Gemini only see stand-ins. The real values come back in the answer, and the audit log shows the route. If your rules don't allow any global model, run Dalla with local models only.
Is this a private ChatGPT for our organization?
Close. Your staff get a chat interface that works like ChatGPT, with your documents, assistants, and automations. It runs in your own cloud tenant, in your data centre, or on our UAE cloud, with per-user keys, budgets, and a full audit trail. Developers can also call an OpenAI-compatible API, so the tools you already have keep working.
Does our data leave the UAE?
Not in local-only mode. Storage, inference, logs, and embeddings stay in the country, and no masking is needed. If you add global models, personal data is masked before anything leaves, so the model only sees stand-ins. The audit log records the route of every request either way.
What is the difference between data residency and inference residency?
Data residency is where files sit at rest. Inference residency is where the GPU actually runs the prompt. A regulator asking where the data went cares about both. With local models, Dalla keeps both in the country. With global models, it masks personal data first, so what is processed abroad carries stand-ins, not the real values. The log shows user, model, cost, and route.
What happens to personal data in a prompt?
When a prompt goes to a global model, Dalla masks it first. Emirates IDs, IBANs, mobile numbers, names, and similar fields are replaced with stand-ins that keep the same format, in English and Arabic, so the model still understands the document. The real values are restored in the answer. Masking is optional: you choose when it runs and which kinds of data it covers. Local-only setups don't need it.
Are our prompts kept or used to train the models?
No training. Your prompts and files aren't used to train any model. How long chats and files are kept is your setting, down to zero data retention (ZDR), where no prompt is stored at all and only the audit record of user, model, cost, and route is kept.
Is a prompt processed abroad and returned still a cross-border transfer?
We can't answer that as legal advice, and we'd be wary of any vendor who does. What we can say is what Dalla does: in local-only mode nothing goes abroad. With global models, only masked text does, the log shows the route, and your compliance team gets the evidence to make the call.
Can Dalla answer from our own documents?
Yes. Upload policies, contracts, and reports to Knowledge, including scanned PDFs, Word, Excel, and PowerPoint. Dalla uses retrieval-augmented generation (RAG) to find the right passages, answers from them, and cites the source. Documents are masked like any other prompt before they reach a global model.
Does Dalla include project management?
Yes. Work is part of Dalla: boards with table, kanban, calendar, timeline, and dashboard views, plus docs, public forms, and approvals. It uses the same sign-in, roles, and projects as the chat, and it's installed wherever Dalla is, so your project data stays inside the same perimeter.
Can we assign tasks to the AI?
Yes. Dalla is in the people list like a colleague. Give it a task and it works with what you can see, comments what it did, and closes the task when it's done. Changes that need approval wait for the approver. It doesn't phone or message anyone outside your organization.
Can clients and contractors use our boards?
Yes, as guests on the boards you share with them. Columns you hide never reach their browser, and they can comment only if you allow it.
Do you support single sign-on?
Yes. Sign-in with Microsoft Entra ID or any SAML provider is set up for your deployment, with users and groups kept in sync. Custom roles, groups, and projects decide who can use which models, tools, and knowledge.
Does Dalla work with AI agents and tools?
Yes. Dalla connects to tools through MCP (Model Context Protocol) servers. An admin connects a server once, and each group of tools is set to on, ask first, or off. Tool inputs are masked like prompts. Agent-to-agent (A2A) support is on the roadmap.
What is sovereign AI?
The models run on infrastructure you can place legally and operationally: in your country, in your own tenant, or on your own servers. Data residency is where the bytes sit. Sovereignty also covers who operates the stack, who can be compelled to hand data over, and whether you can prove the path of a request.
Does Dalla help with PDPL?
Dalla keeps personal data in the country, or masked when a prompt goes to a global model, and gives you the access controls and audit trail a PDPL (Personal Data Protection Law) review asks for. It does not replace your compliance team. It gives them the route of every request and the reports they need to sign off.
Can we run Dalla on our own servers or cloud tenant?
Yes, and most teams do. Dalla installs in your own Azure or other cloud tenant, on-premise, or in an air-gapped network. Contact us and we will scope it with your team.
Does Dalla work for teams outside the UAE?
Yes. Teams in Saudi Arabia, Qatar, Bahrain, Oman, and Kuwait run the same install in their own tenant or data centre, in their country. Teams outside the region are welcome too.
Can our developers keep their existing tools?
Yes. Dalla exposes an OpenAI-compatible API, plus an Anthropic-compatible one, so code and tools built for either work by changing the endpoint and the key.
Which models can we use, and what does it cost?
Local models hosted in the country, models you run on your own GPUs, and, if your policy allows, Claude, GPT, and Gemini with masking. Dalla routes each request to the right one, so you do not pay top price on every call.
We already use ChatGPT Enterprise or Copilot. Where does Dalla fit?
Many teams keep those for general work. Dalla is for work that touches confidential data: it runs in your tenant, works with local and global models, masks personal data, and keeps an audit trail your regulator can read.
We have Azure OpenAI in UAE North. Can Dalla use it?
Yes. Dalla runs in your Azure tenant and uses the model deployments you have approved, next to local models. It adds chat, masking, knowledge, keys, budgets, and the audit log.
Does Dalla work with Core42 Compass?
Yes. Compass gives you models on sovereign UAE compute. Dalla is the workspace on top: chat for staff, masking, knowledge, spend caps, and an audit trail a PDPL review can read.
أسئلة شائعة بالعربية
هل تغادر بياناتنا الإمارات؟
لا، في وضع النماذج المحلية فقط. يبقى التخزين والاستدلال والسجلات داخل الدولة، ولا حاجة لإخفاء البيانات. وإذا أضفتم نماذج عالمية، تُخفى البيانات الشخصية قبل أن يغادر أي شيء، فلا يرى النموذج إلا بدائل. ويسجّل سجل التدقيق مسار كل طلب في الحالتين.
هل هذا ChatGPT خاص لمؤسستنا؟
يحصل موظفوكم على واجهة محادثة تشبه ChatGPT، مع مستنداتكم ومساعديكم وأتمتتكم. تعمل داخل حسابكم السحابي، أو في مركز بياناتكم، أو على سحابتنا في الإمارات، مع مفاتيح وميزانيات لكل مستخدم وسجل تدقيق كامل. ويمكن للمطورين استخدام واجهة برمجية متوافقة مع OpenAI.
هل يمكننا استخدام Claude وGPT وGemini؟
نعم، وبطريقة يوافق عليها فريق الامتثال. تُخفي دلّة البيانات الشخصية داخل الدولة قبل إرسال الطلب إلى نموذج عالمي، فلا يرى إلا بدائل، ثم تُعاد القيم الحقيقية في الإجابة. وإذا كانت قواعدكم لا تسمح بأي نموذج عالمي، تعمل دلّة بالنماذج المحلية فقط.
هل تُستخدم أسئلتنا لتدريب النماذج؟
لا. لا تُستخدم أسئلتكم وملفاتكم لتدريب أي نموذج. مدة الاحتفاظ بالمحادثات والملفات إعداد بأيديكم، حتى عدم الاحتفاظ بأي شيء، فلا يبقى إلا سجل التدقيق: المستخدم والنموذج والتكلفة والمسار.
ماذا يحدث للبيانات الشخصية في الطلب؟
عند إرسال الطلب إلى نموذج عالمي، تُخفى أولاً. تُستبدل أرقام الهوية الإماراتية وأرقام الآيبان وأرقام الهواتف والأسماء ببدائل تحافظ على الصيغة نفسها، ثم تُعاد القيم الحقيقية في الإجابة. الإخفاء اختياري، وتحددون متى يعمل وأي أنواع البيانات يشمل.
هل تجيب دلّة من مستنداتنا؟
نعم. ارفعوا السياسات والعقود والتقارير، بما فيها ملفات PDF الممسوحة ضوئياً وWord وExcel وPowerPoint. تبحث دلّة في المستندات بتقنية التوليد المعزز بالاسترجاع (RAG)، وتجيب منها مع ذكر المصدر.
هل تشمل دلّة إدارة المشاريع؟
نعم. وحدة Work جزء من دلّة: لوحات تُعرض كجدول أو كانبان أو تقويم أو خط زمني أو لوحة مؤشرات، إلى جانب المستندات والنماذج العامة والموافقات. وتستخدم تسجيل الدخول والأدوار والمشاريع نفسها التي تستخدمها المحادثة، وتُثبَّت أينما ثُبّتت دلّة، فتبقى بيانات مشاريعكم داخل النطاق نفسه.
هل يمكن إسناد مهام إلى الذكاء الاصطناعي؟
نعم. تظهر دلّة في قائمة الأشخاص مثل أي زميل. أسندوا إليها مهمة فتعمل بما يراه من أسندها، وتكتب تعليقاً بما أنجزته، وتغلق المهمة عند الانتهاء. والتغييرات التي تحتاج إلى موافقة تنتظر صاحب الموافقة. ولا تتصل بأحد ولا تراسل أحداً من خارج مؤسستكم.
ما الفرق بين إقامة البيانات وإقامة الاستدلال؟
إقامة البيانات تعني مكان تخزين الملفات. إقامة الاستدلال تعني مكان المعالج الذي يشغّل الطلب فعلياً. الجهة الرقابية تسأل عن الاثنين. مع النماذج المحلية تُبقي دلّة الاثنين داخل الدولة، ومع النماذج العالمية تُخفي البيانات الشخصية أولاً. وتسجّل لكل طلب المستخدم والنموذج والتكلفة والمسار.
هل تساعد دلّة في الامتثال لقانون حماية البيانات الشخصية؟
تُبقي دلّة البيانات الشخصية داخل الدولة، أو مخفية عند إرسالها إلى نموذج عالمي، وتوفّر ضوابط الوصول وسجل التدقيق الذي تطلبه مراجعة قانون حماية البيانات الشخصية. لا نحل محل فريق الامتثال لديكم، بل نعطيه مسار كل طلب والتقارير التي يحتاجها للموافقة.
هل يمكن تشغيل دلّة على خوادمنا أو في حسابنا السحابي؟
نعم، وهذا ما تختاره أغلب الجهات. يمكن تثبيت دلّة داخل حسابكم على Azure أو أي سحابة أخرى، أو في مركز بياناتكم، أو في شبكة معزولة. تواصلوا معنا ونحدد النطاق مع فريقكم.
هل تدعم دلّة الدخول الموحد؟
نعم. يُضبط تسجيل الدخول عبر Microsoft Entra ID أو أي مزوّد SAML لكل نشر، مع مزامنة المستخدمين والمجموعات. وتحدد الأدوار والمجموعات والمشاريع من يستخدم أي نموذج وأي أداة وأي مستند.
أي النماذج يمكننا استخدامها؟
نماذج محلية مستضافة داخل الدولة، ونماذج تشغّلونها على معالجاتكم، وإذا سمحت سياستكم، Claude وGPT وGemini مع إخفاء البيانات الشخصية. توجّه دلّة كل طلب إلى النموذج المناسب حتى لا تدفعوا السعر الأعلى على كل استدعاء.
Request access
Get a private walkthrough
Tell us a bit about your team. We'll set up a walkthrough and talk through how Dalla fits your rules, including your own tenant, on-prem, and managed options.
- A 30-minute call with the people who built it
- A written data-flow summary your compliance team can keep
- Pricing for tenant, on-prem, and managed hosting