Operation and data handling
| Aspect | How it is done |
|---|---|
| Operating location | Dedicated infrastructure at Hetzner, data centres in Germany. US hyperscalers are excluded by design because of US jurisdiction. |
| Customer environments | A dedicated, isolated Kubernetes cluster per customer environment. Full kubectl access is available on request — you do not have to operate the cluster yourself. |
| Language model | A locally hosted Mistral model on a dedicated AI server. External models are called only where they are explicitly configured for a task. |
| Model choice | Set per task in the process model, not globally. The choice is versioned and visible in the audit trail. |
| Logging | Input, model, version, result, role and timestamp for every agent step in a case. |
| Process format | BPMN 2.0. Models are exportable and readable without DYNAMIXS.AI. |
| Integration | A universal API adapter to email (IMAP), FTP, SharePoint, AD and LDAP, OpenID Connect, ERP systems such as SAP, Oracle and Salesforce, plus external databases (SQL, S3). |
| Hosting certification | The data centres are certified to ISO/IEC 27001:2022 — Hetzner Online GmbH, scope: all hosting services and data centres. |
Designed for EU AI Act. The platform is built for the requirements of the EU AI Act: role binding, logging, human oversight at defined points and a traceable model choice. That is not the same as a formal conformity assessment, and none is claimed here.
Three layers, clearly separated.
Core
The process engine and identity management. It holds the state of a case, who is responsible and which deadlines apply — and it is the only component that triggers a step.
Services
Administration, retrieval, OCR, document editing and the language models. Each is a task the process can call, never a service that calls the process.
External systems
Reached through one API adapter rather than point-to-point integrations, so a change of source system does not reach into the process model.
What runs in a customer environment.
The platform is a multi-tenant SaaS built on Jakarta EE and operated from container images. Each customer environment runs its own stack:
What an auditor actually asks for.
Rarely the architecture. Almost always a single case. For every agent step the platform holds three things:
Trigger
Case ID, process version, the step that triggered it and the role responsible.
Basis
The data sources and documents used, with model identifier and model version.
Outcome
The agent's output, the approval given or still pending and the timestamp.
The result is a case file in the language of the process, not a model export that somebody has to interpret.
An architecture review with your own team.
We go through the setup with the people responsible for security and operations — technical, without a pitch deck.
Arrange a reviewDYNAMIXS.AI GmbH · Munich · info@dynamixs.ai · +49 89 452136-22