Services
AI engineering
AI is not a business unit next to the others here. It is two things: a capability in the product and a way of working in the team.
When we are the right fit - and when we are not
Where AI carries weight - and where it is merely expensive.
A fit when …
- You have time series or sensor data and want decisions from it, not just charts.
- A language model belongs inside an existing product, not beside it.
- You need to settle where data may be processed before the technology is chosen.
- Your team should work agent-assisted without losing traceability.
Not a fit when …
- There is no reliable data foundation yet - then data engineering is the first step, not the model.
- A rule set demonstrably solves the problem better and more cheaply.
- You expect a model to intervene in safety-relevant ways without human sign-off.
- The goal is a trade fair demonstrator, not production use.
1. AI in the product
Forecasting and time series analytics
Predictive models for electricity price, generation and consumption as the basis for automated load decisions - trained and evaluated on high-resolution time series with Python, Pandas, NumPy and scikit-learn against TimescaleDB.
Anomaly detection and predictive maintenance
Analysis of machine and sensor data for early detection of failures, with Kafka as the basis for real-time event streaming, detection and alerting logic.
Language models in the application
Integration of large language models into product interfaces - as a voice assistant via an STT → LLM → TTS pipeline for natural-language querying of technical systems, or as a context-aware help and analysis function over product-internal data.
Computer vision
Image and document recognition with OpenCV and OCR, among other things for automated verification of identity documents on mobile devices.
2. Agent-assisted development
We use AI agents productively in our own development process - embedded in the existing GitLab toolchain and secured by the same quality gates as hand-written code.
- Orchestrated agent teams - specialised implementation and review agents divide the work; a separate review agent checks every change before it goes to human review.
- Ticket-driven workflow - agents only work on issues that have passed a Definition of Ready. Requirements, acceptance criteria and scope exist before the first commit.
- Anchored in CI/CD - every agent-generated change goes through build, unit, integration and E2E tests plus static analysis. Nothing bypasses the pipeline.
- Tool integration via MCP - connected to repository, issue tracker, documentation and test environments, so context does not have to be supplied manually.
- Automated routine runs - headless operation for dependency updates, test coverage analyses, documentation reconciliation and refactoring proposals.
- Humans stay accountable - architecture, security-relevant matters and final sign-off remain with people. Agents accelerate implementation; they replace neither review nor responsibility.
We pass this experience on as a consulting and training service: introduction into existing teams, tool evaluation, securing the process through quality gates, and data protection and compliance questions when using commercial models.
3. Operations, data sovereignty and model selection
At industrial clients, one question regularly comes before the technology: where may the data be processed?
- Assessment of cloud API, dedicated hosting and full self-hosting - including cost, hardware and energy considerations rather than gut feeling.
- Data protection and confidentiality tiers of commercial providers, training-data exclusion, retention.
- Architecture for the hybrid case: sensitive processing local, uncritical load external.
- Operation and monitoring of inference workloads in your own observability stack.
Technologies
Typical project size
What a project costs with us.
Day rate €480-960 depending on the type of engagement
Below 10 person-days our ramp-up effort does not pay off for you. An assessment of data sovereignty and model selection also works as a short entry point.
Frequently asked questions
Do you use AI agents in client projects as well?
Yes, within defined guardrails. Agents work exclusively on issues that have passed a Definition of Ready; every change goes through build, unit, integration and E2E tests plus static analysis. Architectural decisions, security-relevant matters and sign-off stay with people.
What happens to our source code?
We settle that in writing before the project starts. We assess cloud API, dedicated hosting and full self-hosting, review providers' training-data exclusion and retention periods, and design the hybrid case where needed: sensitive processing local, uncritical load external.
How do we know it actually helped?
By the same metrics as always: throughput, test coverage, defect density after release. Here the purpose of the freed-up capacity is explicitly test depth rather than feature count - and that is measurable.
Contact
Your contact
Software architecture & technical project management