Internal knowledge assistant
Persian Q&A over regulations, contracts, and procedures with the source shown, without document text going to a foreign service.
Private AI and local language models on your own servers
Models and workflows run on your servers or in an Iranian data centre instead of a foreign service. Your documents stay there too, and as long as every part of the system is inside the network, an international outage does not stop work.
What private AI means here
Here the model and application run on your own servers or on a server in an Iranian data centre, and the documents stay there. We use open-weight models, whose published weights run on your hardware. On your servers, data stays in the organisation; in an Iranian data centre, it stays in the country. This service is not resale of access to foreign models.
Common use cases
Summaries and drafts
Summaries of letters, reports, and meeting minutes, and draft replies for a specialist to review and approve.
Document extraction
Key fields extracted from contracts, invoices, and forms, with ambiguous cases routed to a review queue.
Workflows with n8n
n8n on your own server connects email, accounting software, CRM, and messengers to the model, for example to route incoming letters to the right unit.
Replacing a foreign API
Workflows that depend on a foreign model API or an intermediary gateway are moved one by one to a local model so the next outage does not stop them.
Units with sensitive data
Legal, finance, or HR teams that cannot send contracts, financial statements, or personnel files to an outside service.
How we work
The model is chosen after several open-weight candidates are tested on your own Persian tasks, and hardware is sized once the workload is known. Neither is promised in advance.
- 01
Discovery
Which tasks must run locally, what data they involve, how many concurrent users there are, and what response time is enough.
- 02
Model evaluation
Several open-weight models are compared on a versioned set of your real tasks, and each model's licence is checked for commercial use.
- 03
Hardware sizing
GPU needs are estimated from model size, concurrent users, and response time, and the model is tested on existing or rented hardware before anything is bought.
- 04
Build
Model server, Persian retrieval over internal documents (RAG), the assistant interface, and n8n integrations on your infrastructure.
- 05
Pilot and handover
The system is tested on real work and once with international connectivity deliberately cut, then handed over with documentation and team training.
Controls needed from day one
- Access control
- Each user signs in with an organisational account and only gets answers from documents their role may see.
- Logging and audit
- Questions, retrieved documents, answers, and workflow actions are recorded with time and user.
- Network isolation
- The model server has no internet route and only approved internal systems can reach it. If needed, the whole system runs fully offline.
- Updates and patching
- Versions are pinned and documented; each update comes through a controlled internal path and is tested against the evaluation set first.
- Backup
- The document index, configuration, n8n database, and encryption keys are backed up, and restores are tested.
- Evaluation set
- The test examples stay in use after launch, so any drop in quality after a change is caught.
Zharf
Try the ZharfAI assistant right now
Ask about Persian assistants, document automation, or how to start a pilot.
Frequently asked questions about private on-premise AI
Does it work when the international internet is cut?
Yes, if the model, document index, assistant application, n8n, user sign-in, and monitoring are all inside the network and users reach the system over the domestic network. Any component that depends on a foreign service stops during an outage, so the pilot includes a test with international connectivity deliberately cut.
What hardware is needed?
It depends on model size, concurrent users, and response time, and is estimated after discovery. GPUs are expensive and harder to procure in Iran, so we first test on existing hardware or a rented server in an Iranian data centre. For simple tasks, a small model without a GPU is sometimes enough.
Is the quality as good as ChatGPT or Claude?
Not always. Open-weight models that run on an organisation's servers are usually smaller and may be weaker than the largest foreign models on some tasks, though they may be enough for narrow tasks such as answering from documents. The pilot measures this on your real work.
Can it connect to our existing systems?
Yes, through n8n installed on your own server or directly through a system's API or database. Email, accounting software, CRM, and messengers are common examples. The connection method is chosen after reviewing each system, usually starting read-only.
How much does private on-premise AI cost?
It depends on the number of tasks, document volume, model size, hardware, and integrations. After discovery you receive a staged estimate: first a pilot on one task, then expansion. Hardware is estimated separately because it can be the largest cost.
Start with one task that must not stop
In the first conversation we review which task must keep running without international connectivity, its documents, its users, and the hardware you have. If a local model is not good enough, we say so plainly.