I build AI features into the software you already run: language models, retrieval augmented generation and agents that give verifiable answers instead of guessing – as a freelancer, remote across Europe and on-site in North Rhine-Westphalia.
Most companies do not need another chatbot. They need AI exactly where time is lost today: searching through documents, sorting incoming requests, moving data between systems. That is where I start – inside the software you already use, not in a separate tool nobody opens.
As a freelance senior software engineer I bring both sides: the AI integration and the application and cloud engineering underneath it. An AI feature is rarely just an API call. It needs clean data access, permissions, error handling, cost control, monitoring and a deployment that still makes sense three months later.
When data must not leave the building, the model runs locally – through Ollama or LocalAI in your own data centre or in a European cloud region. Where a managed service is enough, I use Azure OpenAI, the Claude API or OpenAI. We make that call up front, because it shapes architecture, cost and data protection.
Four building blocks that work on their own or together – from first prototype to production.
Connecting Claude, OpenAI, Azure OpenAI or local models to your applications and APIs: drafting text, summarising, classifying requests or extracting structured data from free text – including prompt versioning, cost control and a fallback when a model does not answer.
Your company knowledge becomes searchable: documents, wiki pages, tickets or ERP data are prepared, indexed in a vector database and answered through semantic search – with a source for every statement and permission checks per user.
Agents that do more than answer: they call your systems through well-defined tools, triage tickets, validate master data or prepare quotes. The Model Context Protocol (MCP) lets me expose existing interfaces cleanly and reuse them.
On-premise or EU region, pseudonymisation before the model call, logging of answers and human approval before anything goes out. Plus evaluation tests, so a prompt change does not quietly degrade quality.
Scenarios that have proven themselves in practice – and where the benefit is measurable.
People ask in plain language about policies, contracts or schedules and get an answer that links back to the original document. It removes the hunt through scattered drives and shortens onboarding considerably.
The knowledge often sits in the ERP rather than in documents. I build the CDS view or OData service inside SAP myself and connect the language model to it – through tool calls instead of an export. The asking user's authorisations still apply. There is more on this on the SAP integration page.
Invoices, delivery notes, quotes or forms are read and handed to the target system as structured data – with confidence scores and a review screen for the cases that need human approval.
Incoming requests are categorised, prioritised and enriched with matching knowledge articles or draft replies. The team still decides, but no longer spends its time sorting.
Existing web applications gain focused AI features: suggestions while filling in forms, plausibility checks, summaries of long histories or a search that understands synonyms and domain terms.
AI projects rarely fail because of the model. They fail on vague requirements and missing measurement. Hence this sequence.
We define which task the AI should take over, which data it needs and how we measure success. I will say openly when something is cheaper to solve without AI.
A lean prototype on your actual documents, not on sample data. Within days you can see how good the answers are and whether further investment is worth it.
Connecting your systems, permissions and logging, tests against a fixed question set, cost limits and monitoring. After that the feature runs in daily business, not only in a demo.
Deployment through CI/CD, documented architecture and a handover to your team. I stay available for further development and maintenance if you want that.
What is used depends on your data, your data protection requirements and the infrastructure already in place.
Describe briefly where your processes lose time today. I will come back with an honest assessment of whether and how AI helps there.