Work

Four kinds of system, one construction principle.

We build agentic systems, knowledge systems, adapted models and image analysis. They share one structure: the model generates, a check layer decides, every step is logged.

Systems

What we build

  • 01

    Agentic systems

    Multi-step agents with tool access to your existing systems. Output is schema-validated, write actions are tied to approvals, and every action is recorded in an audit log your compliance team can read.

    For the IT team
    Orchestration with LangGraph/LangChain or the Microsoft Agent Framework; models via API (Anthropic, OpenAI, Azure OpenAI) or self-hosted behind an OpenAI-compatible vLLM endpoint — switching is configuration, not a rebuild. Integration with existing systems via REST APIs or direct database access; structured outputs with schema validation, so downstream systems receive reliable data. Guardrails and human-in-the-loop approvals at the points you define; every action lands in an audit-proof log. Delivered as Docker containers, with Kubernetes manifests on request.
  • 02

    Knowledge systems (RAG)

    Hybrid retrieval with sources at document and section level. Search respects your existing access rights, and the hit rate is measured on reference questions from your department — before acceptance and after every update.

    For the IT team
    Retrieval-augmented generation (RAG) with a vector database and hybrid search — semantic plus keyword; chunking and metadata per document type. Permission mapping onto your existing roles and groups: anyone not allowed to open a document will not get it as an answer either. Every answer carries its source at document and section level. An evaluation pipeline with reference questions from your domain — retrieval quality is measured before acceptance and re-checked after every update; an ingest pipeline keeps the index current as documents change.
  • 03

    Model adaptation

    Where standard models do not know your terminology, forms or norms, we adapt open models to your documents. We choose the base model by benchmark on your data and demonstrate the improvement against the agreed criteria before delivery.

    For the IT team
    LoRA/QLoRA fine-tuning of open models with PyTorch; the base model is chosen by benchmark on your data. Training and evaluation data are prepared from your own stock with documented steps; models, datasets and prompts are versioned, every training run reproducible. Delivered quantised for your GPU servers or as a vLLM deployment — with a before/after measurement against the agreed criteria.
  • 04

    Image and sensor data

    Models for image, 3D and measurement data, from classification to segmentation — for in-line quality inspection, for example. Inference runs at the edge or on the GPU server; latency and detection rate are recorded in the acceptance protocol.

    For the IT team
    PyTorch training on GPU and HPC infrastructure (CUDA), from classification and segmentation to 3D volume data. Inference optimised for where it runs — quantised at the edge next to the line, or on a GPU server in the data centre. Integration with control systems and existing software via common interfaces such as REST or OPC UA; latency and detection rates are measured and recorded in the acceptance report.
Structure

How every system is built

How a system is built 01 INPUT 02 AGENT 03 CHECK LAYER 04 APPROVAL 05 EXISTING SYSTEM Inputdocument · requestevent Agentmodel + toolsstructured output Check layerschema · rulesmeasured on criteria Approvala person decideswhere you define it Existing systemERP · CRMdatabase · API finding goes back — correction Audit-Logevery step recorded and traceable MODEL RUNS VIA cloud API EU cloud your own servers
Engineering

What separates a production system from a demo

  • 01

    Evaluation as a regression test

    Reference questions and test cases run before acceptance and again after every change to model, prompt or data.

  • 02

    Structured output

    Models answer into a fixed schema. Whatever fails validation reaches no downstream system.

  • 03

    Permissions, not trust

    What an agent may execute is set by a permission layer, not by the prompt.

  • 04

    Foreign text is not a command

    Content from documents, web pages and tool output is data, never instruction. That is the basis of defence against prompt injection.

  • 05

    Traceable down to the single step

    Input, model version, tool call and result are in the audit log; models, data snapshots and prompts are versioned.

  • 06

    Built model-independent

    Cloud API, EU cloud or your own servers sit behind the same interface. Switching is configuration.

Operations

Three deployment paths

 Cloud APIEU cloud, self-hostedYour own servers
ModelsOpenAI, Anthropic or Microsoft AzureOpen models in a German/EU data centreOpen models on your hardware — up to fully air-gapped environments
Time to startDaysWeeksWeeks, plus hardware procurement
Cost profileUsage-based, no own hardwareMonthly server costs, predictableOne-off hardware, low running costs
Your documentsStay in your systems; only the individual request goes to the provider, encrypted (DPA, processing in EU regions, no training on your data)Stay entirely in the EU data centreNever leave the building
Typical choice forFast start, variable load, standard confidentiality requirementsUS providers ruled out, predictable costsTrade secrets, highly confidential data, regulated environments
After go-live

Measure, update, hand over

  • 01

    Monitoring and drift

    In operation, answer quality is checked against the reference measurement.

  • 02

    Updates with a regression test

    An update only goes live when the measured values hold.

  • 03

    Handover without dependency

    Source code, documentation and deployment live in your repositories.

How a project runs is described under How we work.