Hafez Fabric

Governed AI Gateway

Put AI to work across your organization without sending your confidential data outside it. Fabric runs on your own infrastructure, anonymizes sensitive data before it reaches a cloud model, and governs every model, knowledge base and agent.

Request Demo Contact Us Datasheet (PDF)

About Fabric

Most organizations are caught between two bad options. Cloud models are capable and make work faster, but sending a contract, a customer file or a security report to a service outside the organization is not acceptable, and is sometimes not legal. The usual outcome is one of two things: the tools are banned and the advantage is lost, or people quietly paste the data into personal accounts and nobody knows what left.

Fabric removes the trade-off. It is self-hosted on your own infrastructure and sits between your people and the models: it substitutes sensitive values before a request leaves and restores them when the answer returns, keeps each team and client separated, builds answers from your own documents with citations, runs agents only with registered tools and behind human approval, and records every request.

HafezSecure built this platform for its own work, because our security teams face the same limit when handling client data. We use it every day, and we now offer it to organizations that have the same problem.

Anonymized

Sensitive Data on the Way Out

Self-Hosted

On Your Infrastructure

Isolated

Workspace per Team and Client

Audited

Every Request Recorded

Your Data Never Leaves the Organization

For organizations worried about confidentiality when using cloud models: Fabric substitutes sensitive data before the request is sent, and puts it back once the answer arrives

Sensitive Values Are Recognized

Before a request leaves your network, Fabric reads it, including the text in attached images and PDFs, and finds the confidential values. Iranian national IDs, card and Sheba numbers, phone numbers, email addresses, IP addresses and internal hostnames are recognized automatically, with the Iranian identifiers’ checksums verified. Names of people and clients, addresses, contract numbers and project codes are covered by the lists and patterns your administrators define for each workspace.

They Are Replaced, Then the Request Is Sent

Each value is swapped for a labelled placeholder such as [[IR_NATIONAL_ID_1]], and the same value always gets the same placeholder, so the question still makes sense: one customer stays one customer throughout the text. Only this anonymized version travels to the external model. The provider receives a coherent question and none of your data.

The Answer Comes Back Complete

The answer arrives carrying the placeholders. Inside your own network, Fabric puts the real values back where they belong and hands the finished answer to your colleague. The reader sees a normal, complete answer; the confidential parts of it never crossed your boundary.

What Gets Replaced

National IDs, card and Sheba numbers, recognized automatically
Phone numbers and email addresses, recognized automatically
Internal hostnames and IP addresses, recognized automatically
Names of people, clients and suppliers, from your lists
Addresses, contract numbers and project codes, from your patterns
Text inside images and PDFs, read with Persian OCR

What Stays Intact

The question itself and what it asks for
The structure and meaning of the text
The relationship between values, kept consistent
The quality of the answer you get back

Once enabled for a workspace, the same rule applies in chat, in knowledge bases, in agent runs and in application calls. For the most sensitive work, restrict a workspace to local models and no request leaves the organization at all.

Two Layers of Protection

Patterns catch the values; a local AI model understands the documents. Both run inside your network.

Layer 1: Rules Before Anything Leaves

Once Persian text is normalized and text inside images and PDFs is read with OCR, national IDs, card and IBAN numbers are recognized by their check digits, alongside your own lists and patterns. For each kind of value the administrator decides what happens: record it, mask it, or block the request. With protection on, images and files are not sent to an external model unless the administrator allows it.

Layer 2: AI That Reads for Meaning

Patterns cannot tell a confidential letter from a public notice; a model running inside your network can. It reads uploaded documents, and outgoing prompts if you choose, and labels them with the categories your organization defines, such as confidential letter, board minutes or contract, following instructions you write in plain language. By default it runs only on a local model, so nothing leaves to be judged, and with no permitted model it does not run at all. Every label, never the text, is written to the audit log, so you can see which confidential documents are being used with AI.

Labels are recorded for review; masking and blocking stay with the first layer.

One Platform, Three Layers

Risky execution is kept away from control and from data, so a mistake in one layer does not reach the next

The Control Layer

Sign-in, workspaces, permissions, model routing, anonymization rules, quotas and the audit trail. This is where an administrator decides what is allowed, and where every request is recorded.

The Knowledge Layer

Your documents, how they are indexed, and how answers are built from them with citations. Document permissions follow through to retrieval, so a person cannot reach through the assistant to read what they could not open directly.

The Work Layer

Where models run and agents do their work, limited to the tools you register and a restricted outbound path. Risky execution is kept here, away from the portal and the data: agent tools run behind a deny-list, and outbound traffic goes through an egress broker with an allow-list.

How Fabric Is Built

Four views of the architecture: what happens to a request, how teams are kept apart, where the platform runs, and how answers and actions stay traceable.

Before a request reaches any model, Fabric checks who is asking, whether that model is allowed for them and whether quota and budget remain; then it masks what must not leave and records the cost in toman.

A Look at the Portal

The portal your team actually uses — Workload, Knowledge and Governance in one place: governed chat with citations, masking before anything leaves, and a record of every request

Figures and entries shown are an illustrative sample, not live production data.

Fabric — Governed AI Gateway

Preview

Overview — every workspace, one governed gateway

Search workspaces, models, keys…

Last 30 daysAll providersAll workspaces

Workspaces

Isolated at the database level

41k

+18%

Requests this month

Chat, knowledge bases, agents, keys

48,310

+6,120

Values masked

Across 13.9k outbound requests

62%

Of quota used

18 days remaining

The path of every request, this month

41k
Requests
→
37.7k
Passed policy
8% refused
→
13.9k
Masked
Bound for external
→
37.6k
Answered
Local or external
→
37.6k
Restored and recorded
Every one audited

Requests per day

4 weeks ago1,520 todayToday

Where the requests go

Local models58

External, masked34

Refused by policy8

Usage by workspace

Assessment delivery78% of cap

Client — bank A44% of cap

Internal knowledge26% of cap

Eight per cent of requests were refused outright by policy — a model a workspace may not reach, or a key bound to a different knowledge base.

An illustration of the interface with sample data. Switch screens with the tabs above.

Key Features

What makes organizational AI trustworthy and accountable

Anonymization Before the Cloud

Use the strongest external models without handing them your confidential data. Sensitive values are masked on the way out and restored on the way back, and, once masking is on for a workspace, the activity log shows what was masked in every request.

One Gateway for Every Model

Models you run in-house and models you buy sit behind one compatible endpoint. An application written for a common AI interface works unchanged, and switching provider is an administrative setting rather than a development project.

Knowledge Bases That Cite Their Sources

Upload your policies, contracts, reports and manuals, and get answers built from them with a reference to the passage behind each claim. Each person retrieves only from documents they are allowed to open, so the assistant never becomes a way around permissions.

Agents That Ask Permission

An agent can read a report, query a system and prepare a result, but only through tools you registered, and it stops for human approval before a step that changes anything. The person who requests a run is not the person who approves it.

Separation Between Teams and Clients

Workspaces are separated in the database itself, not only by a filter in the application, and access is decided per member. A consulting engagement, a subsidiary or a department can share one installation without sharing a single document.

Every Request on the Record

Who asked, which model answered, which documents were used, what an application call cost and what left the network: all of it is written down. Retention is a policy you set, so logs live as long as your rules require and no longer.

Keys Bound to a Purpose

When another application calls the platform, its key can be tied to one knowledge base and one prompt template with its own quota. A leaked key is then worth very little: it opens one narrow door, and its use shows up in the same audit trail.

Nothing Leaves Without Permission

Outbound traffic goes through an egress broker with an allow-list, so a model, an agent or a connector reaches only destinations you approved; everything else is refused and recorded. Each approved destination is held to the address it was checked at, so it cannot be redirected to an internal system mid-request.

Planned: Connected to Your Security Work

Connectors to the Hafez Security Platform are planned: they would bring findings and asset data into the same workspace, so an analyst could ask about an exposure and get an answer grounded in the organization’s own current data. The platform is not deployed yet, so they are not available today.

Quotas That Hold

Each member and application key gets a request and token quota, so one team or one integration cannot use up the platform. Calls through the API are priced in toman as they happen, so you see what each application costs.

See How an Answer Was Produced

Each answer can be opened up: which model was used, which passages were retrieved, which tools ran and how long each step took. When a result looks wrong, you can check it instead of arguing about it.

Files in Chat, Without Leaving a Trail

A colleague can attach a document to a conversation and ask about it. The file belongs to that session, follows the same rules as everything else, and does not quietly become part of a shared knowledge base.

A Persian Interface, Right to Left

The portal works in Persian, with a right-to-left layout and Jalali dates, as well as in English, with light and dark themes. Staff use it in their own language from the first day.

What AI Costs, in Toman

A model-prices page compares hundreds of models from domestic and international gateways side by side, priced in toman, and the gateway can pick the cheapest model your rules allow for each call. Choosing a model becomes a cost decision you can see.

Documents and Images, Not Just Text

Text inside attached images and PDFs, Persian included, is read before anything is sent, and a classifier running on a local model can flag documents such as confidential letters so your rules treat them accordingly.

Use Cases

Wherever AI work meets data that must not leave

Banks and Regulated Industries

Customer records, transactions and contracts stay inside the bank while staff still use capable models for drafting, review and analysis.

Security Teams Working on Client Data

Findings, logs and reports from one engagement are readable only inside that engagement, and client names never travel to an external provider.

Organization-Wide Assistants

One assistant over your policies, procedures and manuals, answering with citations, so staff stop guessing at internal rules.

Automating Security Workflows

Triage, enrichment and first-draft reporting run as governed agent tasks, with a person approving anything that leaves a mark.

Legal, Audit and Compliance Review

Long contracts and audit evidence are summarized and compared without the documents themselves leaving the organization.

Sectors With Legal Confidentiality Duties

Health, government and other fields where sending personal records to an outside service is simply not permitted.

How It Works

From installation to everyday use

Install It on Your Infrastructure

Fabric runs on servers you control, inside your own network. Our team installs it with you, connects it to your sign-in system, and hands over an environment your administrators own from the first day.

Connect Your Models

Register the models you want: ones running on your own hardware, ones from a provider you already buy from, or both. Then decide which teams may reach which model, and which work must stay on local models only.

Write the Rules Once

Define what counts as sensitive, who may use what, how much each member and application may use, how long logs are kept, and which agent steps need a human. These are settings in the portal, not code, and they apply everywhere at once.

Your Teams Get to Work

People use chat, knowledge bases and agents through one portal, and other applications call the same gateway. Every request is governed, anonymized where your rules require it, and recorded.

Deployment Options

From a single server to a network with no route outside

A Single Server

Enough for a department or a first rollout: the whole platform on one machine, with your own database and storage.

A ClusterPlanned

In development: for heavier use, the same platform will run across a cluster you operate, keeping teams and workloads separated as they grow.

Designed for Offline Networks

Designed to run with local models only and no route to the internet at all, for classified networks and air-gapped sites.

Why Fabric

How it differs from using AI services directly

Confidentiality Enforced by the Platform

The question "can we use this model without leaking our data?" is answered by the platform, not by asking every colleague to be careful. Once an administrator turns it on for a workspace, the rule is enforced on the way out, for everyone, on every path.

You Own the Whole Platform

It is self-hosted on your own hardware and designed to run without any outside service. For organizations that cannot depend on foreign providers, that is the difference between adopting AI and postponing it.

Governance Instead of Guesswork

Access, quotas, retention and approvals are administered centrally, so use can spread across the organization without anyone losing sight of the cost or the risk.

Built by a Security Company, In Daily Use

Fabric is the platform HafezSecure built for its own work and runs every day. Isolation, auditing and least privilege are part of its design rather than features added after an incident.