AI that runs inside your walls.
We design, deploy and run private AI on your own servers: language models, search over your documents, analytics, and agents for internal work. Then we train your team to run it.
Ask a question. Get the answer and the page it came from.
This is what staff see. Pick an office, tap a question, then open the source. Every answer points to the passage it was taken from, so a person can check it in seconds.
Library · 4 documents
Tap a question
Rent is ₮18,000,000 per month, payable by the 5th. Late payment accrues a penalty of 0.5% of the overdue amount per day. Rent is indexed by 5% on each anniversary of the start date. Either party may terminate with three months' written notice after the first 24 months. The tenant may not sublet without the landlord's written consent.
Example documents and answers, written for this page. On your own system the answers come from a model running on your hardware, reading your files.
For organizations whose data can't leave the building.
Banks, utilities and government bodies in Mongolia and Central Asia that can't send documents or customer data to foreign clouds.
- Language models deployed on your hardware
- Search and answers over your internal documents
- A data warehouse and analytics layer
- AI agents for internal workflows
- Training and handover to your team
- A record of every question, kept on your servers
For places where the data is the business.
Banks and lenders
Customer files, credit memos and policies stay inside the bank.
- Search policies, procedures and past decisions
- Draft replies and reports from your templates
- Answer questions from the data warehouse, with the query shown
Utilities and infrastructure
Metering, billing and network data, read without touching the control systems.
- Find unusual consumption and rank what to inspect
- Forecast load by feeder or district
- An operator assistant that explains the numbers in Mongolian
Government bodies
Citizen and case records that should not sit on a foreign server.
- Sort and route applications and complaints
- Search regulations, decisions and archives
- Read forms and scans into your systems
Law firms, clinics and accounting offices
One office, a few dozen people, and files your clients trust you to keep.
- Ask your contracts, records and procedures
- Find the clause, the rate or the deadline in seconds
- A box on your shelf instead of a server room
The everyday work of your staff, faster.
Ask your documents
Staff ask a question in plain language and get an answer with links to the documents it came from.
Draft from your own templates
Letters, reports and replies drafted from your templates and past documents, for a person to check and send.
Sort and route requests
Incoming requests, complaints and emails sorted by topic and sent to the right team.
Read forms and scans
Fields pulled out of PDFs, scans and forms and entered into your systems.
Answer from your data
Questions about your own numbers answered from the data warehouse, with the query shown so it can be checked.
Agents for routine work
Multi-step internal tasks such as looking up, filling in and checking, with a person approving the result.
The same system reads your operational data.
Find what is unusual
Meters, transactions or records that stop behaving like the rest, ranked into a worklist so inspectors start with the likeliest.
Forecast what is coming
Short-term forecasts of load, demand or cash, checked against what actually happened.
Read drone and site images
Defects and changes found in imagery and linked to the asset record they belong to.
One set of numbers
Data from your separate systems joined in a warehouse, with each metric defined once so every report agrees.
Score the risk
Each asset or account gets a risk score that a person can trace back to the readings behind it.
Keep models honest
Models are re-tested as new data arrives, and replaced only when the new one does better.
Eight rules. They go in the contract.
These are not slogans. We write them into the statement of work.
- 01Nothing crosses the line without a signature.
Everything runs on your hardware. If a path in or out is ever needed, such as a bank feed or weather data, it is written down, drawn on the diagram and approved by you first.
- 02It can look, not touch.
Where control systems are involved, data flows one way: from operations to the AI. No part of the system has a path back into them.
- 03The model explains. It does not decide.
Losses, anomalies and financial figures are computed by logic you can read and audit. The language model describes the result in plain words. It never produces the number.
- 04An answer comes with its source, or not at all.
Every answer cites the passage it came from. When the documents don't hold the answer, the assistant says so instead of guessing.
- 05The model never holds the keys.
It has no passwords and no direct access to your systems. It goes through a gate with three levels: read, a business action, and an irreversible action, which always waits for a person. Every request is written to a log that cannot be edited.
- 06One number, one formula, one source.
Each metric is defined once, tested and documented, so the dashboard, the report and the assistant all give the same figure.
- 07Your team can run it without us.
Training and documentation in Mongolian are deliverables, not extras. By the end your staff operate, monitor and extend the system.
- 08Accepted in writing, piece by piece.
Each module has acceptance criteria agreed before work on it starts. You sign off when it meets them, not when the calendar says so.
Control a cloud service can't give you.
Your data stays put
Documents, questions and answers stay on your network. Nothing goes to an outside AI service.
You decide who sees what
You set who can use it and which documents each team's assistant can read.
No fee per question
You pay for the hardware and our work, not for every question your staff ask.
Three ways to bring AI into your organization.
| A cloud AI service | Building it yourself | With PrimaRoas | |
|---|---|---|---|
| Where your documents go | To the provider's servers, outside your network | They stay on your network | They stay on your network |
| Works without internet | No | Yes | Yes |
| Who chooses the model | The provider | You | You, with our tests on your documents |
| What you pay for | Every seat or every question | Hardware, and a team you have to hire | Hardware and a fixed fee |
| Who keeps it running | The provider | Your own AI engineers | Us first, then your staff, trained by us |
| Quality in Mongolian | Varies by service | Depends on your team's testing | Tested on your own documents before roll-out |
| The record of who asked what | Held by the provider | Only if you build it | Built in, on your servers |
Everything sits inside your network.
The whole system runs on servers you own or in a data centre you choose. Each layer can be inspected, and none of them needs an outside service to work.
Outside AI services: not used
What your security team will ask, answered.
Bring your security and IT people to the first call. These are the points they usually raise.
- No internet requiredThe system runs on an isolated network. It does not call home.
- Access by teamEach team's assistant reads only the folders and data that team is allowed to see, using your existing staff accounts.
- A log of everythingEvery question, answer and action is recorded on your servers, ready for your compliance team.
- Updates you installNew models and software arrive as signed bundles. You decide when they go in. Nothing updates itself.
- Approval before actionThe assistant reads freely within its permissions. Anything that changes a record waits for a person.
- No lock-inOpen-weight models, and code you own. If you stop working with us, the system keeps running.
It works with what you already have.
During roll-out we connect the system to your existing files and software. Nothing has to be replaced first.
- Shared folders and document systems
- PDFs, scans and forms
- Email and incoming requests
- Databases and data warehouses
- ERP, billing and finance systems
- Metering and SCADA data, read-only
- ArcGIS and drone imagery
- Your staff accounts
Three ways to start.
Sanaa Box
A box on your shelf for a law firm, clinic or accounting office. It reads your contracts, records and files and answers questions about them.
- We install it and index your existing files
- Staff training included
- New files in your shared folder are read automatically
- One price for the box, one yearly fee for support
A programme
Several use cases on your own servers, delivered in stages, with your team learning to run it.
- Assessment, pilot, roll-out, handover
- Connected to your systems and staff accounts
- Acceptance criteria agreed before each stage
- A monthly fee during delivery, a lower one for support
Licence the stack
You sell and install in your market. We supply the software, the updates and second-line support.
- Document search and chat in local languages
- Updates as signed offline bundles
- An audit log for compliance teams
- A console under your own brand
Sanaa Box sizes
We confirm the size with you before we quote. The price covers setup, indexing your existing files and staff training.
| Size | For | Hardware |
|---|---|---|
| S | Up to 25 people, about 50,000 documents | 1 GPU, 2 TB of storage |
| M | Up to 60 people, faster on long documents | 2 GPUs, 8 TB of storage |
| L | Firms with branches | A rack server with a standby, so one failure doesn't stop work |
Start small, prove it, then grow.
- 1Assessment
We look at your data, systems, security rules and hardware, and choose the first use case with you. You get a written report: what your data can support, the hardware it needs, and a goal for the pilot.
- 2Pilot
One use case running on your servers, used by real staff and measured against the goal agreed up front. You decide whether to go on with the results in front of you.
- 3Roll-out
More use cases and more teams, connected to your systems and your staff accounts. Each module is accepted in writing against criteria set before it started.
- 4Handover and support
Training and documentation so your team runs it. Then support from us for a lower monthly fee: fixes, re-testing models as your data changes, and a review every quarter.
You sign off on results, not on effort.
The tests are written down before the work starts. These are the ones we use.
- Reports and dashboardsEvery figure traces to one definition and matches the source system over an agreed test period.
- Detection and forecasting modelsEach flag points to a specific meter, asset or account. Accuracy meets the level agreed at the start, on data the model has not seen.
- The assistantIt answers in Mongolian with its sources, declines when the data is not there, and every action that needs approval waits for one.
- HandoverYour staff have been trained, hold the documentation in Mongolian, and have run the system themselves under our supervision.
What people ask us.
Do we need to buy hardware?
Usually one or more servers with GPUs, sized to how many people use the system and which models it runs. We size it during the assessment. It can also run on hardware you already have, or in a data centre you choose.
Which AI models do you use?
Open-weight language models that run entirely on your hardware, chosen for each task and tested on your own documents. You aren't tied to one AI vendor, and models can be swapped as better ones come out.
Can it run without an internet connection?
Yes. Once it's installed, the system can run on a network that isn't connected to the internet.
How well does it work in Mongolian?
Language models differ a lot in Mongolian. We test candidate models on your own documents during the assessment, so you see the quality before you commit to a roll-out.
Who owns what you build?
You do: the deployment, its settings and the code we write for you, with documentation and training so your team can run it.
Can it take actions in our systems, or only answer?
Both, within limits you set. Reading is open within each team's permissions. A business action, such as creating a work order, goes through a gate with rules. Anything irreversible always waits for a person's approval, and every action is logged.
What if the AI gets something wrong?
Answers show the passage they came from, so a person can check them, and the assistant is built to say when the documents don't hold the answer. For figures and decisions, the result comes from logic you can audit, not from the language model.
How is this different from giving staff ChatGPT?
A cloud assistant doesn't know your documents unless you upload them, and uploading sends them to someone else's servers. This runs on your network, reads your own files under your access rules, and leaves the record with you.
What does our IT team need to do?
Give us a place for the server, network access inside your building, and one person to work with us. We do the installation and the connections, and train your team to take over.
How do updates work without internet?
We deliver new models and software as signed bundles, on a drive or over a controlled connection. Your IT team installs them when you choose.
What happens if we stop working with you?
The system keeps running. The deployment, its settings and the code we wrote for you are yours, and the models are open-weight, so nothing switches off.
How much does it cost?
A Sanaa Box is one price for the box plus a yearly support fee, quoted once we know how many people will use it. A programme is a monthly fee during delivery and a lower one for support. Either way you pay for hardware and our work, not for each question.
How long does it take?
It depends on the scope. The assessment is the shortest stage and ends with a plan and dates for the pilot. We agree the timeline in writing before the pilot starts.
Priced by the box or by the programme.
A Sanaa Box: one price and a yearly support fee. A programme: a monthly fee while we deliver, then a lower one for support. A licence: a yearly fee and a share of support revenue. We agree the scope and the goal with you before anything starts.
Tell us what you're building.
We reply within one business day. Prefer email? Write to contact@primaroas.com.