Kauzas AI Platform
Your assistant cannot see data its user cannot see.
Kauzas AI Platform builds search, assistant and agent capabilities over your organization's data. Identity and permission awareness is not a feature added later; it is the foundation the product is built on. It installs on top of your existing data infrastructure — whichever one that is.
- On any data infrastructure
- Model independent
- Runs on a closed network
kauzas ai
GenAI services · one front end
- Front end
- Agent runtime
- Cedar
- MCP
- Vector + search
- SearchKeyword + meaning
- AgentsLangGraph
- PermissionsKeycloak · Cedar
- ToolsMCP servers
- ModelsProvider independent
- Skill packsPer customer
- GuardrailsPrompt injection defence
- ObservabilityLangfuse · budgets
Where it installs
Where your data sits is not this product's problem.
Kauzas AI Platform runs stand-alone. Data from your different systems — ERP, CRM, IoT, documents — is brought together on a data mesh; the platform reaches that mesh through a single connection and can query all of it at once. What sits underneath that mesh is your decision.
Kauzas Data Platform
In-house · AWS Local Zone Istanbul
AWS
S3 · Glue · SageMaker Unified Studio
Microsoft Azure
Fabric · OneLake · Synapse
Google Cloud
BigQuery · Cloud Storage
Databricks
Unity Catalog · Delta Lake
Your own data lake
Existing, or built in-house
Changing data infrastructure is a decision measured in years and budgets. Starting with enterprise AI should not have to wait for it: you begin with the infrastructure you have, and if that infrastructure changes later the platform stays where it is.
Identity and permissions
Enterprise AI projects do not stall for technical reasons.
The pilot works, the demo lands well, and then one question stops it: does this assistant know who is allowed to see which data? If the answer is no, the project goes no further — because an assistant that lets everyone see everything is the fastest route to a data leak an organization has.
The usual arrangement
- 01The assistant connects to the data source under its own technical identity.
- 02That identity is broadly privileged; it has to serve every user.
- 03The model reads data the user is not permitted to see.
- 04The user's permissions are applied only at the interface, after the answer is produced.
Control has been left to the model's output.
Kauzas AI Platform
- 01The user signs in through Keycloak.
- 02Cedar policies decide which agents and which tools they can reach.
- 03The tool connects to the data source under that user's identity.
- 04Data outside their permissions never reaches the model.
Control is applied before data reaches the model.
If a user cannot see a piece of data, neither can the assistant working on their behalf — and it cannot talk about data it cannot see.
Capabilities
Gathered into one front end.
Search, agents, skill packs, model settings and the setup of MCP servers all sit in the same interface. Setting up an MCP server is itself subject to the authorization rules; opening up the interface does not mean loosening control.
Hybrid search
Both by keyword and by meaning and similarity. A user who does not know what something is called inside the organization still reaches the right result.
LangGraph agents
Not single-turn question and answer; workflows made of several steps that decide as they go.
Model independence
It works with whichever model or provider you want. Models in this field turn over within months.
Skill packs
Separate packs per customer or sector. The same platform is the same product in a bank and in a plant; what differs is the skills laid on top.
Prompt management and A/B testing
Version management over skill packs and prompts, so the claim that a change improved things becomes measurable.
Evaluation layer
An eval layer measuring the quality of agent and model output is available and switched on when needed.
Freshness
When data in the mesh is updated, that freshness carries through to search and indexing. The assistant does not answer with yesterday's picture.
Feedback loop
A user marking an answer wrong feeds back into the system, so quality stops being something set up once.
Multilingual
No separate language switch. The models are already multilingual; a user writing in any language gets an answer in it.
Guardrails
A protection layer against harmful and unwanted output and against attacks such as prompt injection.
Disconnected networks
It can be installed and operated in air-gapped environments with no internet connection.
MCP servers
Agents connect to MCP servers running in the background with specific permissions. Setting them up is subject to the same rules.
Oversight
AI usage is not a space outside the organization's own oversight.
Opening an assistant to the organization should not mean giving up sight of what is being asked and what is being spent.
Who asked what
Every query and conversation is held in Langfuse. Who asked what can be seen and audited.
Who can spend how much
Usage limits and budgets can be defined per user or per group. In a period when model costs are hard to predict, that is an operational necessity.
The same permission at both ends
The user's identity applies both to which tools they can reach and to what those tools see in the data source. Permissions need not be defined twice.
Frequently asked questions
- Do I need Kauzas Data Platform to use Kauzas AI Platform?
- No. The platform runs stand-alone and works with existing data infrastructure such as your own lakehouse, Microsoft Fabric or Amazon SageMaker Unified Studio. Kauzas Data Platform is only one of those substrates.
- Can the assistant see data the user is not permitted to see?
- No. The user's identity is carried through both to the tools they can reach and to the data sources those tools connect to. Data the user cannot see never reaches the model.
- Which AI models can be used?
- The platform is model and provider independent; it can work with whichever model or provider the organization prefers.
- Does it run in environments with no internet connection?
- Yes. The platform can be installed and operated in air-gapped environments.
- Which data sources can it connect to?
- Data from different systems such as ERP, CRM, IoT and documents is brought together on the data mesh; the platform reaches that mesh through a single connection and can query all of it at once.
- Is multilingual use supported?
- Yes, without a separate language switch. The models are multilingual, so a user writing in any language gets an answer in that language.
- Can AI usage be audited?
- All queries and conversations are held in Langfuse, so who asked what can be seen. Usage limits and budgets can also be defined per user and per group.
- How is output quality measured?
- An evaluation layer that measures the quality of agent and model output is available and can be switched on when needed. Skill packs and prompts support version management and A/B testing.
- Is there protection against harmful output?
- A guardrail layer protects against harmful or unwanted output and against situations such as prompt injection.
Let's see it on your own data.
We can show how an assistant would be built on your existing data infrastructure, working inside your own permission model.