AI software development for media companies: from idea to production system
We build the AI applications that live on after the proof of concept: integrated into your systems, operated GDPR-compliantly, and measured by results.
The real work lies between demo and daily newsroom use.
A chat prototype is built in an afternoon. A system that runs reliably at six every morning is engineering – and that is what we do.
Grown system landscapes
An editorial system from the 2000s, an archive in three formats, a CMS with home-grown plugins: we know this reality and integrate AI without you having to migrate everything first.
Scaling after the PoC
What worked with 50 test articles collapses at 50,000 – technically and financially. We dimension pipelines, costs and quality assurance for real throughput.
Veto power of IT security & legal
Many AI projects die in the approval process, not in engineering. We deliver the answers your reviewers need: data flows, DPAs, hosting evidence, deletion concepts.
What we build.
Prototyping & MVP
From prioritised use case to a clickable system with real data in 4–6 weeks – so your newsroom tests before your company invests at scale.
LLM & RAG applications
Language models working on your content: retrieval augmented generation against hallucinations, with source citations and editorial control.
CMS & editorial system integration
AI where the work happens: as a function inside your editorial system, not as yet another browser tab. We integrate via APIs, plugins and middleware.
NLP & ML models
Tagging, entity recognition, classification, semantic search and recommendation logic – as robust pipelines, not one-off scripts.
MLOps & operations
Monitoring of quality and cost, evaluation, model updates, fallbacks. An AI system is never “done” – we operate it so it gets better, not worse.
API development
Clean interfaces so AI functions are equally usable in app, website and internal tools – documented and versioned.
Proven tools, replaceably assembled.
We build model-agnostic: your application should be able to switch providers when prices, quality or regulation demand it.
Typical building blocks
- Python and TypeScript for services and integrations
- Commercial LLMs via EU endpoints or open-source models (e.g. via vLLM)
- Vector databases and hybrid search for RAG
- Evaluation pipelines with editorial quality criteria
- Container deployment in EU cloud, private cloud or on-premises
How we work
- Iteratively in short cycles, with your newsroom as test instance
- Measurable acceptance criteria instead of “looks good”
- Documentation and hand-over capability from day one
- Cost monitoring as part of the architecture
- One accountable contact from discovery to operations – see approach
Security and compliance are architecture, not an appendix.
This section is for the colleagues who have to sign at the end. The short version: you will not get a black box from us.
Data location & hosting
EU hosting as the default, on-premises or private cloud for sensitive assets. Data flows are fully documented – including what a model provider sees and what it does not.
DPA & GDPR
Data processing agreements under Art. 28 GDPR with us and all subprocessors, input for your records of processing activities and deletion concepts are part of the delivery.
Privacy by design
Data minimisation, pseudonymisation and purpose limitation are architecture decisions we take before the first line of code – not retrofits for the audit.
EU AI Act
Risk classification, transparency duties under Art. 50 (applicable since August 2026), machine-readable labelling of AI-generated content and documentation for your compliance file.
In daily use: AI in our own portal.
Our portal with 500,000+ monthly users is our own production environment: AI supports research and editing, tags content for search and SEO, and safeguards quality via automated checks – always with human final approval before publication. This system teaches us what models really deliver in daily operations, what they cost and where they fail. That operating experience flows into every client project.
Frequently asked questions about AI software development
How fast can we have a first working prototype?
Typically after 4 to 6 weeks – with real content from your company, not demo data. The prototype runs in a protected environment and is tested by your newsroom. Then you decide about scaling based on measured results.
Which models and providers do you use?
The use case decides, not a preference: commercial models via EU endpoints, European providers, or open-source models in your own data centre. We build systems so the model remains replaceable – prices and quality change faster than your software should.
Can our data leave Germany or the EU?
Only if you explicitly want that. By default we work with EU hosting and data processing agreements; for sensitive content we offer on-premises or private-cloud operation. We document data flows so your data protection officer can audit them.
How do you prevent hallucinations in editorial workflows?
Through architecture, not hope: RAG with source binding, confidence thresholds, automated checking rules and mandatory editorial approval wherever content gets published. With us, AI delivers drafts and suggestions – the publication decision stays human.
What does running an AI system cost?
That depends on volume and model choice – which is why we build cost monitoring in from the start and calculate business cases with realistic token and infrastructure costs. Many editorial use cases land in the low four-digit monthly range, often below.
Do you take over existing PoCs from other vendors?
Yes. We usually start with a review: what is solid, what needs rebuilding? Then you decide whether we continue or replace selectively. A failed PoC is not wasted money if its lessons flow into the second iteration.
Still unclear where to start?
If the use case is not yet settled, development is the second step. Our AI consulting first prioritises where AI has the biggest lever in your company – with business case and roadmap, before budget flows into software.
Let's discuss your project.
Bring your use case – or your stuck PoC. We will tell you honestly what it needs to become productive.