On-prem GPU · Private cloud GPU

Own your intelligence instead of renting it.

Private AI inference on hardware you control.

The AI model runs on a GPU in your office or your own cloud account. On top sits one company-wide AI agent system, and each person sees only what they're allowed to see.

  • Open-weight models
  • Per-person permissions
  • Human approval
How narmar.ai is designed Inside one company: its documents flow into a GPU machine that runs the AI model, with a custom agent system on top. Answers go back to the company's own people, and each person sees only the files they're allowed to see. Your company Office or own cloud account Your documents Contracts · policies · scans CUSTOM AGENT SYSTEM MODEL open-weight GPU machine Runs the AI model Your people Leadership Approves outputs Staff Answers + sources Operations Only allowed files Documents to the machine Answers to people allowed to see them How narmar.ai is designed Inside one company: its documents flow into a GPU machine that runs the AI model, with a custom agent system on top. Answers go back to the company's own people, and each person sees only the files they're allowed to see. Your company Office or own cloud account Your documents Contracts · policies · scans CUSTOM AGENT SYSTEM MODEL open-weight GPU machine Runs the AI model Your people Leadership Approves outputs Staff Answers + sources Operations Only allowed files Documents to the machine Answers to people allowed to see them
How it works, illustrated.

Your people want AI. Your documents can't go just anywhere.

Staff already use AI chat tools for drafting and research. Many companies and practices hold documents that clients, contracts, or their own rules keep out of outside services. So the usual choices are to block AI or to look away. narmar.ai is built to give you a third choice: own your AI and run it yourself, instead of renting it from an outside AI service.

What people want

  • Help drafting and summarizing
  • Answers from the company's own files
  • Less time on repeated paperwork

What the company has to protect

  • Client and patient documents
  • Terms in contracts about where files may go
  • Its own rules about who may see what

Two places the model can run

A typical cloud AI tool Common today

Questions and documents travel to the AI company's servers to be processed.

With a cloud AI tool, documents leave the company and are processed on an outside service's servers. Your company Outside AI service PROCESSED HERE With a cloud AI tool, documents leave the company and are processed on an outside service's servers. Your company PROCESSED HERE Outside AI service

Private AI inference How narmar.ai is designed

The model runs on a machine inside your company, so questions and documents are processed there.

With private AI inference, documents go to a machine inside the company and are processed there. Your company Office or own cloud MODEL PROCESSED HERE With private AI inference, documents go to a machine inside the company and are processed there. Your company Office or own cloud MODEL PROCESSED HERE

A simplified picture of where the model runs. It describes the design, not a guarantee.

Two layers: the machine, and the system on it.

Every option below has the same two layers. Only where the machine lives changes.

Layer 1 is the GPU machine that runs an open-weight AI model. Layer 2 is the agent system on top: company Q&A, personal agents, workflows, permissions, approvals and an activity log. LAYER 2 · THE AGENT SYSTEM Company Q&A Personal agents Workflows Permissions Approvals Activity log Set up for the way your company works LAYER 1 · THE MACHINE OPEN-WEIGHT MODEL runs on the GPU Layer 1 is the GPU machine that runs an open-weight AI model. Layer 2 is the agent system on top: company Q&A, personal agents, workflows, permissions, approvals and an activity log. LAYER 2 · THE AGENT SYSTEM Company Q&A Personal agents Workflows Permissions Approvals Activity log Set up for the way your company works LAYER 1 · THE MACHINE OPEN-WEIGHT MODEL runs on the GPU
The two layers, on one machine.
Layer 1

The machine

An open-weight AI model runs on a dedicated GPU machine, in your office or in a cloud account your company owns. Questions and documents go to that machine, not to an outside AI company.

Layer 2

The agent system

Assistants and agents, step-by-step workflows, permissions, approvals, and an activity log, set up for the way your company already works. Agents act only within each person's permissions, and nothing goes out until a person approves it.

Getting started

Three steps

  1. We pick one repeated workflow with you and study it together.
  2. The system is set up on your hardware.
  3. Your people use it, and a person approves anything that goes out.

One agent system for everyone in the company.

It comes with every option, set up for your company.

Included

Company assistant

Answers questions from your company's own documents and policies. It's read-only.

Included

Personal assistant

One for each person, set up for their work. Its settings can only take abilities away, never add them.

Included

Document workflows

Step-by-step review of repeated documents, starting with NDAs. Rules decide each finding, and the AI's quotes are checked against the document.

Included

Research drafts

Memos built from your own files, labeled as drafts for a person to review.

Included

Files, scans, and audio

Read in and searchable under the same access rules as everything else.

In development

Agent builder

Describe a repeated task in plain words, and it drafts an agent for it.

Per deployment

Connections to your other tools

Connections to email, calendar, document systems, and Microsoft 365 are built for each deployment and scoped in the assessment.

Three ways to run it.

Pick where the machine lives. Each option includes the same custom agent system, and the hardware is chosen for your company.

A GPU machine inside your office receives your documents and answers your people. Your office GPU machine CUSTOM AGENT SYSTEM A GPU machine inside your office receives your documents and answers your people. Your office GPU machine CUSTOM AGENT SYSTEM
Option 1

On-prem GPU

A private GPU machine in your office. The AI model runs on hardware your company owns, and questions and documents go to that machine.

In your office
Includes the custom, company-wide agent system.
Your office connects to dedicated GPU servers rented in your company's own cloud account. Your office Your cloud account Rented by your company Dedicated GPUs CUSTOM AGENT SYSTEM Your office connects to dedicated GPU servers rented in your company's own cloud account. Your office Your cloud account Rented by your company Dedicated GPUs CUSTOM AGENT SYSTEM
Option 2

Private cloud GPU

Dedicated GPUs rented in your own cloud account, with a provider such as AWS. The machine isn't in your office, but it runs in your company's account rather than a shared AI service.

In your cloud account
Includes the custom, company-wide agent system.
Research idea: the spare memory of three machines you already own is pooled into one memory space that can hold a larger model. RESEARCH · EARLY ACCESS Spare memory on machines you already own One pooled memory space for a larger model CUSTOM AGENT SYSTEM Research idea: the spare memory of three machines you already own is pooled into one memory space that can hold a larger model. RESEARCH · EARLY ACCESS Spare memory on machinesyou already own One pooled memory spacefor a larger model CUSTOM AGENT SYSTEM
Option 3 · Research

RAM pooling

Running larger models by pooling the memory of machines you already own, so together they can hold a model too large for any one of them.

It's an active research track rather than a product, open to early-access participants who want to try it with us.

Research · Early access
Designed to include the same custom agent system.

What every option includes

A custom, company-wide AI agent system that runs on the same hardware as the model:

  • Agents that work with your documents, and with your other tools once those connections are built.
  • Each person's agents see only what that person is already allowed to see.
  • Nothing goes out until a person approves it. No agent can approve its own work.
  • An activity log of who did what, without storing document text.

For anyone whose documents can't go just anywhere.

It was built first for law firms, and it's designed for any company or practice with confidential documents.

  • Accounting and tax

    Answers from your own files and procedures.

  • Wealth and financial advisory

    Help with repeated document work.

  • Insurance

    Summaries of long policies and files.

  • Engineering and architecture

    Search across project documents.

  • HR and people teams

    Policy questions, under strict access rules.

  • Any team with confidential files

    Tell us your workflow, and we'll say honestly whether it fits.

Built first for legal work

For law firms.

narmar.ai was built first for law firms. NDA review runs step by step: rules decide each finding, the AI's quotes are checked against the contract, and anything uncertain goes to a person. Search follows matter access and ethical walls. Research memos are drafts for an attorney to review. Nothing goes out until someone approves it.

  • NDA reviewA step-by-step workflow where code, not the model, decides each clause's status.
  • Matter access and ethical wallsA walled document is blocked wherever it's filed, even for people whose group could otherwise see it.
  • Approvals before anything goes outAgents can only ask. A person decides.
  • Research drafts for attorney reviewBuilt only from the firm's own documents, and labeled as drafts.

How NDA review runs

  1. The contract comes inRead under the same access rules as every other document.
  2. The model pulls out factsFor each clause, with the exact quote it relied on.
  3. Code checks every quoteThe quote must be in the contract and match the value, or the clause goes to a person.
  4. Rules decide each clauseA missed keyword means "needs review," never "OK."
  5. A person reviews and approvesNothing is delivered until they do.

Step 5 is always a person.

Practice operations

For medical practices and clinics.

The same system is designed for clinics and medical practices, where many documents shouldn't go to an outside AI service.

It's for administrative documents, not patient records or clinical decisions, and it gives no medical advice. Whether a tool fits your obligations is your practice's call.

What it's designed to help with

Notes

Turning a staff member's own rough notes into a clean draft for them to check.

Summaries

Short summaries of long documents, such as guidelines, manuals, and contracts.

Internal policy questions

Answers about your practice's own policies and procedures, with the source shown.

Paperwork

Drafts of routine forms and letters for staff to review before anything is sent.

Who sees what, and where it runs.

Own your data. narmar.ai is designed so your documents, prompts and outputs are processed and kept on hardware you control, in your office or in your company's own cloud account, not by an outside AI service.

  • Each person sees only the documents they're already allowed to see.
  • Ethical walls block a document wherever it's filed.
  • Answers point to their sources. Anything uncertain goes to a person.
  • Nothing goes out until a person approves it.
  • An activity log records who did what, without storing document text.

Where it runs

  • On a GPU machine in your office, or on dedicated GPUs rented in your company's own cloud account.
  • Open-weight models, chosen to fit your work and your hardware.

How it's set up

  • Permissions and walls are set up to match the groups and access rules your company already has.
  • It starts with one repeated workflow, studied with you in the assessment.

Start with an assessment.

We study one piece of repeated work with you, measure how long it takes now, and list the data rules your company says apply. Then we tell you honestly whether to run AI on your own hardware, use a cloud tool you may already pay for, or do nothing new.

We don't ask for client names, matter details, patient information, or client documents.

The assessment
What you get
A map of how the work is done today, how long it takes, the data rules your company says apply, and a written recommendation. If the answer is "don't buy anything," that's what it will say.
Who does the work
We do, mostly by hand.
After it
If running AI on your own hardware makes sense, the system is quoted for your company.

Ask us about scope and pricing

Questions people ask first.

What do you mean by "private AI inference"?

The AI model runs on a machine your company owns, or rents in its own cloud account, so questions and documents are processed there instead of by an outside AI service. It's a description of where the model runs, not a guarantee.

Does it really run on our own hardware?

Yes. The model and the agent system run on a GPU machine in your office or on dedicated GPUs in your company's own cloud account, sized for your work during the assessment.

Which of the three options should we pick?

It depends on your company, and the assessment is where that gets worked out. An on-prem GPU machine sits in your office. A private cloud GPU is rented in your own cloud account. RAM pooling is a research track with early access, for people who want to try it with us.

What is RAM pooling, and can we use it now?

It joins the memory of several computers you already own so together they can hold a larger AI model. It's an active research track rather than a product. If you'd like early access, tick it in the form.

Which AI models does it use?

Open-weight models that run on the machine itself, chosen per company.

Is it secure? Is it compliant?

We don't make those claims. It doesn't include encryption or single sign-on, so it isn't set up to hold real client or patient data. Whether a tool fits your obligations is your company's call.

Why not just use ChatGPT or Claude? Aren't they compliant?

For many companies, they may be enough. OpenAI and Anthropic offer business plans with real controls: business data isn't used for training by default, SOC 2 reports, retention settings, and for some customers zero data retention or a HIPAA business associate agreement.

But you're renting. Your prompts and documents are processed on servers run by the provider or its cloud partner, under their terms, and they set retention options and when models change. Some client agreements, protective orders, regulators and internal policies limit or bar sending certain documents to an outside AI service.

narmar.ai is designed so you own the intelligence and your data: the model runs on hardware you control. Whether any tool meets your obligations is a question for your counsel. We don't make compliance claims.

Can it connect to our email, documents, and Microsoft 365?

Those connections are built for each deployment, and the assessment settles which ones your company needs.

Does it replace ChatGPT or our current tools?

Maybe not. The assessment may well conclude that a tool you already pay for is enough.

What does it cost?

It depends on the company, the work, and the hardware. Ask us about scope and pricing.

Do you record calls?

No. No recordings and no AI note-takers. Our notes are by hand.

Ask for more information.

Tell us a little about your company, and we'll reply within two business days.

We're building this product, so a conversation could lead to a paid assessment.

  • Prefer email? mta@narmar.ai
  • Interested in RAM pooling research? Tick it in the form.
  • Please don't include client names, matter details, patient information, or anything confidential.
This form isn't connected yet, so it can't send anything. Please email mta@narmar.ai instead.

Interested in (optional)

Please don't include client names, matter details, patient information, or anything confidential. Up to 1,000 characters.