Skip to content

Platform · ESTAYA AI Platform

One platform for all your AI applications

The ESTAYA AI Platform is the central layer between your applications and AI models: LiteLLM gateway, citation-grounded RAG, AI agents and governance in one place, on our own servers in Germany. In production, among others, in the AI archive of a major German regional newspaper.

Shadow AI creeps in when every team picks its own tools. The platform puts models, knowledge and agents behind one API, with roles, audit log and budget limits.

CAPABILITIES

What the platform does

Models, knowledge, agents, under control

Multi-model gateway (LiteLLM)

One OpenAI-compatible API in front of many models, open or your own. Routing, failover, load balancing and cost tracking per team and key. Swapping models takes no code change.

RAG & grounding

Cited answers with PageIndex, NLI grounding & knowledge graph.

AI agents

Plan-execute-reflect, tool calling, MCP.

Guardrails

PII protection, topic filters, prompt-injection defence.

Governance

Roles, quotas, allowlists, budgets.

Observability

Quality, latency, tokens and cost become measurable. Audit logs cover what the EU AI Act and internal audit ask for.

Authentication & SSO

Our own identity service (Keycloak) with single sign-on, OIDC/SAML, 2FA, roles and tenants. One login covers every AI service.

Training & reinforcement learning

Fine-tuning, LoRA and RLHF/DPO pipelines on our own GPUs. Feedback from your teams flows back into a measurably better model.

UI framework

Ready-made components for chat, search, dossiers and dashboards. Your AI interface is up and running within days.

App framework & SDK

SDK, APIs and building blocks to ship your own AI apps fast. Auth, RAG and agents come already built in.

Why a platform

Model independence instead of vendor lock-in

Your applications speak a single, OpenAI-compatible format. Which model answers behind it, Llama, Mistral, Qwen, Teuken or your own fine-tune, is decided by the platform according to your rules. Changing models is as simple as changing electricity providers, with nothing to rebuild.

  • One API (LiteLLM) for open and custom models
  • Cost and quota control per team
  • Deploy as SaaS, VPC or air-gapped on-premise
  • Inference with vLLM on our own GPU servers in Germany
Understand sovereign AI
AI Platform

Our own algorithms

More than a RAG toolkit

Beneath the platform sit methods generic toolkits don't ship: hybrid search with Reciprocal Rank Fusion, cross-encoder reranking, NLI sentence grounding with abstention (zero hallucination), an NLI-verified knowledge graph with path-finding and communities, plus context-adaptive weighting by question intent.

Quality is measurable: an eval harness with RAGAS-style metrics (faithfulness, answer relevancy, context precision), a gold set and a regression gate. Confidence is reported honestly (high/medium/low) instead of pseudo-percentages.

  • Reciprocal Rank Fusion + cross-encoder reranking
  • NLI grounding & abstention (zero hallucination)
  • Knowledge graph: paths, communities, brokerage
  • Measurable quality via eval harness (RAGAS)
AI Platform
ONBOARDING

In production within a few weeks

How we roll out the platform

01

Assessment

  • Use cases & data sources
  • Security & compliance scope
02

Setup

  • Set up gateway, RAG & models
  • Roles, budgets, guardrails
03

Pilot

  • First use case in production
  • Evaluation & fine-tuning
04

Scale

  • More teams & use cases
  • Operations, monitoring, SLA

Frequently asked

Questions about the platform

What is a multi-model gateway?

A unified API in front of many models. Your applications speak one format (OpenAI-compatible); the gateway routes, balances load, fails over and tracks cost. Swapping models needs no code change.

Does the platform run fully on-premise?

Yes. Gateway, RAG and inference (vLLM) run entirely in your data centre or air-gapped, with no data egress.

What governance does the platform provide?

Audit logs, budget and quota limits per team and key, model allowlists, roles and guardrails (PII, topic filters), documentable for the EU AI Act and internal audit.

One platform instead of scattered tools

Separate AI components carry a hidden cost: a vector store in one place, an API adapter in another, a script for evaluation somewhere else. The ESTAYA AI Platform brings these layers into one documentable system. At its centre sits a gateway built on LiteLLM. It standardises access to open-weight models such as Llama, Mistral, Qwen and Teuken, which we run ourselves with vLLM. You stay free of vendor lock-in and can swap models without rewriting your applications.

Answer quality comes from several methods rather than a single retrieval step. Citation-grounded RAG ties every statement to its source. PageIndex places long documents in structure, a knowledge graph connects entities, and an NLI check discards claims the context does not support. Our own algorithms sharpen selection: Reciprocal Rank Fusion merges multiple result lists, cross-encoder reranking orders by relevance, and context-adaptive weighting adjusts to the query at hand.

AI agents follow a Plan-Execute-Reflect pattern and connect to your systems through the Model Context Protocol. Governance and observability record every call, model and source so each result can be traced. With the eval_harness and RAGAS we put a number on retrieval quality and faithfulness.

  • Model independence through a single gateway
  • Verifiable answers via grounding and NLI checks
  • Proprietary ranking algorithms beyond plain vector search
  • Deployment on-premises, in your VPC, or as SaaS on our own servers in Germany
  • In production in the ESTAYA AI Platform and a major German regional newspaper AI archive

GDPR, the EU AI Act, BSI IT-Grundschutz and ISO 27001/42001 form the framework we implement for you. Beyonetix holds no certification of its own. What we provide are traceable logs your own audit can use to evidence these requirements.

See the ESTAYA AI Platform on your own data

We set up a pilot in your environment, built around your use case.