Private AI
AI that runs
inside your boundary.
We deploy agents, models and evals where your rules say they must run: an in-country cloud, your own tenancy or your own servers. Your keys, your logs, your sign-off.
Where it runs
The data decides the boundary.
Most rules say where data must live, not which machine it runs on. Classification sets the tier: most regulated work fits a certified in-country cloud, some needs your own servers, and a few sites run with no internet at all. We help you decide, then build to it.
Deployment options
Choose the boundary. We build inside it.
In-country cloud
Certified UAE cloud provider
Processing and storage stay in the UAE on a certified provider. This fits most regulated work.
- Models
- Models hosted in the UAE, or open-weight models
- Data
- Stored in the UAE
- Inference
- In the UAE, on the models offered there
- Keys
- Yours, in a UAE key service
- Trade-off
- Only the models offered in the UAE; the newest can lag
- Hosted models
- In-country only
- Our access
- Approved remote access
- Open route
- Only with your approval
- No route
Illustrative architecture. Each deployment is designed with your security team and recorded in writing.
What we deliver
The work above the infrastructure.
01
Agents and approvals
Workflows built into your systems, with sign-off wherever your rules require it and every action on the record.
02
Model choice and evals
We test candidate models on your own tasks, in Arabic and English, inside the boundary. When a better model arrives, the same evals tell you whether to switch.
03
Tuning, when the evals call for it
Fine-tuning or distillation only when the evals show a gap that prompting and retrieval can’t close.
04
Residency evidence
A written record of where each part runs, who holds the keys and which rules it maps to.
Who does what
Clear lines, in writing.
We don’t resell hardware or run data centres. We work with the provider you choose.
- Servers, GPUs and data centre
- Your infrastructure partner
- Cloud platform and licences
- Your provider
- Workflows, agents and integrations
- HiddenLever
- Model choice, evals and tuning
- HiddenLever
- Keys, identity and data classification
- You
- Approvals and final decisions
- Your named approvers
How it runs
From hosting decision to live workflow.
01
Decide the tier
Classify the data with your security team and decide where each part runs.
Output
Hosting decision, in writing
02
Prove it on your data
Run candidate models against your own tasks inside the boundary, in Arabic and English.
Output
Eval report and model choice
03
Build, with evidence
Agents, approvals and the audit trail, plus the evidence your reviewers ask for.
Output
Residency and keys evidence pack
04
Run or hand over
Updates, model refreshes and drift checks, or a runbook and a trained owner on your side.
Output
Runbook and named owner
Rules we map to
Built to the rules you work under.
We map each deployment to the rules that apply and give your reviewers the evidence. Accreditation decisions stay with your regulator and your security team.
Federal government
Data classification sets the hosting tier. Certified in-country cloud is allowed for most classes.
Abu Dhabi health
Patient data stays in the UAE, the entity holds the keys, and support comes from inside the UAE.
Banks and insurers
A material change to a critical operation needs notice to the Central Bank and an outside expert’s report.
Dubai government
Systems run with certified cloud providers or in the entity’s own data centre.
Questions
Before you decide.
What’s the difference between data and inference residency?
Data residency is where your data is stored. Inference residency is where the model runs on it. Some hosted models keep data in the UAE but run inference elsewhere, so we state both for every part of the system.
Does on-premises mean air-gapped?
No. On-premises means your own servers. Air-gapped means no internet route at all, which needs its own design for updates, models and support. We scope it separately.
Can a frontier model run on our servers?
Hosted frontier models run in their providers’ clouds. On your servers we use open-weight models, including Arabic-capable ones, and test them on your tasks first.
Do you supply hardware?
No. Your cloud or infrastructure partner supplies and runs it. We size what the workflow needs and build what runs on it.
Who holds the keys?
You do. The evidence pack records where every key lives.
Will you fine-tune a model?
Only when the evals show a gap that prompting and retrieval can’t close, usually for narrow, high-volume work. It runs inside the same boundary.
Contact
Start with one workflow.
Tell us where it has to run. We’ll come back in writing with the tier, the model options and what it takes.
hello@hiddenlever.ai