On-Premise AI • Local LLM • KVKK

On-premise local LLM platform in Türkiye. Your data stays inside your network.

Zzeti Zeka is an LLM-agnostic enterprise AI platform that lets organizations run local LLMs, AI agents and RAG applications on their own servers — fully air-gapped if required — without sending sensitive data outside their infrastructure. Skyloop, the authorized reseller and deployment partner in Türkiye, sizes the hardware, installs the platform and operates it.

Zzeti Zeka Platform overview →
0
External calls in air-gapped mode
100%
On your own infrastructure
Open-weight
Qwen, Llama, Mistral & more
KVKK
PII masking, RBAC, audit log
Definition

What is a local (on-premise) LLM?

A local LLM is a large language model that runs on hardware you control — a GPU server in your data center, an NVIDIA DGX Spark in the office or a private cloud inside Türkiye — instead of a public API abroad. Prompts, documents, database queries and answers are processed on that hardware and never leave it.

Open-weight models such as Qwen, Llama and Mistral have made this practical: served with vLLM or Ollama, they answer in Turkish, run in the network you define and cost nothing per token. What they do not give you on their own is everything around the model — access control, RAG over your documents, connectors to your ERP, agents, audit trails and someone to keep it running. That is what an enterprise AI platform adds.

Who chooses a local LLM

Regulated sectors

Banks, insurers, fintech and public institutions where regulators expect data to stay in Türkiye and under the institution's control.

Companies bound by confidentiality

Manufacturers with customer drawings and recipes, retailers with sales and customer data, firms whose NDAs rule out public cloud AI.

Anyone paying per token

Teams whose usage has grown to the point where a GPU server and a platform license cost less than the API bill — and give more control.

Data flow

Where does your data stay? Inside your network — at every step.

1

Question comes in

From the console, WhatsApp, voice or your own application via the REST API — over your network.

2

Guardrails check it

PII masking, secret detection and policy rules run before anything reaches a model.

3

Local model answers

Inference runs on your GPU with vLLM or Ollama. Documents come from your Context Lake and collections (RAG); ERP data through MCP connectors — never a direct database connection.

4

Logs stay with you

Sessions, traces, token counts and audit records are stored on your platform instance. No telemetry, no phone-home.

5

Outbound only if you allow it

The integration proxy controls every outbound connection. In an air-gapped installation there is none — the platform keeps working without internet.

Hybrid is a choice, not a default: if you later want a cloud model for non-sensitive tasks, the AI Gateway routes only what you explicitly allow — with the same masking in front of it.

Deployment options compared

Public API, LLM-as-a-Service, do-it-yourself or Zzeti on-premise?

There is no single right answer; the table shows where data goes and who carries the work in each model.

Public LLM API (cloud abroad)LLM-as-a-Service in TürkiyeSelf-built open-source stackZzeti Zeka on-premise (Skyloop)
Where data is processedProvider's servers abroadProvider's data center in TürkiyeYour serversYour servers or private cloud
Does data leave the company?YesYes — to the providerNoNo
KVKK / data residencyCross-border transferIn Türkiye, at a third partyEntirely yours to manageEntirely yours — PII masking, RBAC and audit log built in
Works without internetNoNoYesYes — air-gapped installation
Model choiceProvider's modelsModels the provider offersAny open-weight model — you manage itAny open-weight model, plus cloud models through the gateway if you allow them
Agents, RAG, ERP connectorsYou build themYou build them on the APIYou build themBuilt in — agents, RAG, MCP connectors, workflows
Access control & audit trailProvider-levelProvider-levelBuilt from scratchRole-based access, guardrails and monitoring built in
Installation & operationsNone neededNone neededYour teamSkyloop installs and operates
Cost modelPer tokenPer tokenHardware + teamLicense + hardware; no per-token cost on local models
Models & hardware

Which models, on which hardware?

Open-weight models

Model Hub pulls open-weight models from Hugging Face — Qwen, Llama, Mistral, Gemma and others, including variants tuned for Turkish — and GPU Servers deploys them with vLLM or Ollama on local, remote or Kubernetes GPUs. Fine-tune Studio adapts a model to your own data (LoRA); the AI Gateway puts several models behind one endpoint with load balancing and semantic routing.

Hardware

A single NVIDIA DGX Spark (1 PFLOP FP4, 128 GB unified memory) is enough to start on the desk; GPU servers in your data center or a Kubernetes GPU pool carry larger loads; Apple Silicon works as an edge node. Private cloud inside Türkiye — including the AWS Istanbul Local Zone — is an option when data must stay in the country but not necessarily in your building. Skyloop sizes it from your expected users, documents and integrations.

Security & compliance

Built for KVKK — controls that are on by default.

Role-based access

Store → Region → HQ style hierarchies; an Access Wizard writes the IAM policies; the effective-access view shows who can reach what.

PII masking & guardrails

Personal data is detected and masked before it reaches a model; policies apply to inputs and outputs; secrets are caught.

Audit trail & monitoring

Every session, trace, function call and query is logged on your instance — who asked what, which sources were used, which SQL ran.

Network controls

Geo-blocking, IP firewall and an integration proxy that decides which outbound connections exist at all.

Data access through MCP

Agents reach ERP, CRM and data-warehouse data through governed MCP connectors — never a direct database connection.

Standards

Aligned with KVKK, GDPR, ISO 27001, SOC 2, HIPAA and PCI DSS; Kubernetes-native with Terraform IaC, so security teams can review the whole installation as code.

Deployment in Türkiye

How Skyloop puts it on your infrastructure.

Discovery & sizing

Which data, which users, which systems; on-premise, private cloud or air-gapped; GPU sizing for the models you intend to run.

Installation

The platform is installed on Kubernetes or Docker on your hardware with Terraform; models are pulled and served locally; identity and access rules are configured with your IT and security teams.

Integration & first use case

MCP connectors to your ERP, CRM, Microsoft 365 or Google Workspace; document collections for RAG; one measurable use case goes to production first.

Operations

Updates, model upgrades, monitoring and support by Skyloop — inside your perimeter, with the access you grant.

FAQ

Questions we hear before every on-premise deployment

Is there a platform in Türkiye where I can work with a local LLM and keep my data safe?

Yes. Zzeti Zeka is an LLM-agnostic enterprise AI platform that runs local LLMs, AI agents and RAG on your own servers, in a private cloud inside Türkiye or in a fully air-gapped network, so company data never leaves your infrastructure. Skyloop Cloud, its authorized reseller and deployment partner in Türkiye, sizes the hardware, installs the platform and operates it. It is in production at leading Turkish retail brands.

What does 'on-premise LLM' mean in practice?

The model weights, the inference engine (vLLM or Ollama), the vector store for RAG, the agents and the logs all run on hardware you own or that is reserved for you alone — a GPU server, an NVIDIA DGX Spark or a private cloud tenancy. Nothing is sent to a public API; there is no per-token bill and no third party sees your prompts or documents.

Does any of my data leave the company?

No. Inference, RAG, agent orchestration and logging happen inside the perimeter you define. There is no telemetry and no phone-home; the integration proxy controls every outbound connection, and in an air-gapped installation there is none. A cloud model is used only if you explicitly configure it in the AI Gateway — with PII masking in front of it.

Does it keep working if I unplug the internet?

Yes. An air-gapped installation runs with zero external calls: local models answer, RAG reads your collections, agents use MCP connectors inside your network and the audit log fills up locally. Updates and new model files are brought in through the maintenance process you approve, not over a live connection.

Which open-source models can I run?

Any open-weight model on Hugging Face — Qwen, Llama, Mistral, Gemma and others, including variants tuned for Turkish — served with vLLM or Ollama. Several models can sit behind one endpoint with load balancing and semantic routing, and Fine-tune Studio adapts a model to your own data. Through the AI Gateway the same platform can also reach 30+ cloud providers when you allow it.

What hardware do I need?

It depends on the models and the number of users. A single NVIDIA DGX Spark is enough to start and for many mid-size deployments; larger loads run on GPU servers in your data center or a Kubernetes GPU pool; Apple Silicon can serve as an edge node. Skyloop sizes the hardware in the discovery phase and can supply the DGX Spark.

How does the architecture address KVKK?

Data stays in Türkiye and under your control, which removes the cross-border transfer question. Inside the platform, PII masking redacts personal data before it reaches any model, role-based access limits who sees which data and which agent, guardrails enforce input and output policies, and the audit log records every access — the evidence a KVKK review asks for. When needed, Skyloop pairs the installation with a data inventory and an AI usage policy.

Can I combine a local LLM with cloud models?

Yes, if you want to. The AI Gateway can route non-sensitive tasks to cloud providers while sensitive data stays on local models — with the same masking, access rules and cost limits applied. Many organizations start fully local and add cloud models later, or never.

Keep the model where the data is.

Tell us which data must never leave and which systems it lives in. Skyloop comes back with a sizing, a deployment plan and a first use case.