Technology

Under the hood.

A good AI system is more than a model with a chat window. It combines language understanding, your knowledge, your tools and clear rules into something that works reliably every day. These are the building blocks.

Building blocks

Language models

The core: models that understand and generate language. We choose them by task, quality, cost and privacy requirements – from large cloud models to compact models that can also run on your own infrastructure.

Knowledge retrieval (RAG)

Instead of answering from memory, the system first searches your own documents and bases its answer on them – traceable, with sources.

Agents & tools

Agents plan steps and use tools to carry them out: search, databases, email or your business applications – connected via APIs and open standards such as the Model Context Protocol (MCP).

Workflows

Clearly defined workflows combine AI steps with classic automation. What works with rules stays rule-based; AI steps in where understanding is needed.

Multimodal

Text, images, audio and documents in one system: reading scans, describing images, transcribing speech or turning content into audio.

Measuring quality

We test with realistic examples from your daily work before a system goes live, and keep an eye on how it performs afterwards – so improvements are measured, not guessed.

PrincipleModel-agnostic

The AI market moves fast. Today’s best model may be outdated tomorrow. That’s why we build so that models can be swapped without redeveloping the whole system.

Which model is used and where it runs, we decide together with you – depending on the task, the budget and your privacy and confidentiality requirements.

SecuritySecurity is part of the build

  • Least privilege: Agents only get the access they need for their task.
  • Approvals: Actions with consequences – sending, paying, deleting – go through a person.
  • Known risks in view: such as manipulated inputs (prompt injection) or unintentionally disclosed information.
  • Logs: What a system has done remains traceable.

GlossaryKey terms, briefly explained

LLM
Large language model – a model trained on vast amounts of text. The foundation of modern AI assistants.
Prompt
The instruction given to a language model. Good prompts are precise, include context and examples – and are maintained like code.
RAG
Retrieval-augmented generation: the model receives relevant excerpts from your documents and answers on that basis.
Embedding
A numerical representation of a text’s meaning. It lets search find content that matches in meaning – even without the exact keywords.
AI agent
A system that breaks a goal into steps, uses tools and checks intermediate results, rather than delivering a single answer.
MCP
Model Context Protocol – an open standard for connecting AI applications to tools and data sources.
Hallucination
An answer that sounds plausible but is wrong. Knowledge retrieval, sources and approvals reduce this risk.
Human in the loop
A person reviews and approves before the AI does something important. Our default wherever it matters.