services / ai integration
AI integration for companies – LLMs, RAG and AI agents in your software
I integrate AI where it creates measurable value: in your existing applications, processes and data. From connecting large language models to RAG over your own documents to AI agents that complete tasks on their own – secure, traceable and GDPR-aware.
starting point
Typical starting points
- Management wants "something with AI", but there is no clear, measurable use case.
- A quick ChatGPT prototype impressed everyone – but in the real application its answers are unreliable.
- Knowledge is buried in manuals, tickets and wikis, and nobody finds the right page in time.
- Recurring tasks such as classifying, summarising or extracting take up hours of your team's time every day.
- Data protection, cost and accountability are unclear, so nobody dares to go to production.
deliverables
What I actually do
- Discovery workshop – prioritise use cases, assess value and risk
- Prototype against real data with clear success criteria
- Integration of LLM APIs (Claude, OpenAI) into your existing systems
- RAG pipelines – document preparation, vector search, citations
- AI agents with tool use and MCP servers for your internal systems
- Evaluation, monitoring, cost and latency optimisation
- Architecture and data protection concept, team enablement
AI that holds up in daily work
A chatbot demo is quick to build. An AI feature that your team and customers can rely on every day is an engineering task. That’s where I come in: I combine experience in software development and software architecture with hands-on work on large language models. The result is AI integrations that behave like normal, well-built software – testable, observable, secure and maintainable.
The focus isn’t the technology but your process. We start by asking where time, money or quality is lost today, and then check whether and how AI can help. Sometimes the answer is an agent, sometimes a simple classifier – and sometimes no AI at all.
From use case to production system
1. Discovery. In a compact workshop we collect use cases, assess value, feasibility and risk, and pick a candidate that can demonstrate value quickly.
2. Prototype with success criteria. I build a focused prototype against real data. Up front we agree how we’ll recognise success – answer quality, time saved or error rate. You decide based on numbers.
3. Integration. If the prototype convinces, I integrate the solution cleanly into your existing landscape: with authentication, permissions, logging, error handling and a UI that suits your users.
4. Operation and evolution. Evaluations run automatically, cost and latency are monitored, and your team understands how the system works and how to extend it.
Typical solutions
- Knowledge assistant with RAG: employees ask questions in natural language and get answers from manuals, contracts or tickets – with citations.
- Document processing: incoming emails, forms or PDFs are classified, summarised and transferred into your systems in structured form.
- AI agents: agents that work with your systems through defined tools – pre-qualifying tickets, doing research or preparing reports. The Model Context Protocol (MCP) provides a standard way to connect internal systems.
- AI features in your product: intelligent search, writing suggestions or assistant features directly in your web application.
Using AI responsibly
Trust is the precondition for using AI in production. That’s why I clarify data flows before the first line of code: which data leaves your company, to which provider, in which region and under which contract? Where it makes sense I use EU hosting and providers that don’t train on your data. I keep the requirements of the EU AI Act in view, document decisions transparently and make sure people stay in control where it matters.
stack
Technologies
- Core
- Strong
- Familiar
AI integration
9
- AI agents & tool use – Core
- LLM APIs (Claude, OpenAI) – Core
- Model Context Protocol (MCP) – Core
- Retrieval-Augmented Generation (RAG) – Core
- Evals & quality measurement – Strong
- Cost & latency optimisation – Strong
- Vector search (pgvector) – Strong
- Fine-tuning – Familiar
- Self-hosted / on-prem models – Familiar
Web & product
8
- Next.js – Core
- Node.js – Core
- React – Core
- TypeScript – Core
- REST & GraphQL API design – Strong
- Astro – Strong
- Testing (Vitest, Playwright) – Strong
- React Native – Familiar
process
How we work together
Arrive on-site first, then deliver efficiently remote – anywhere in Germany and the DACH region.
- 01
Intro call
30 minutes, free – goals, scope and whether we're a good fit.
- 02
On-site onboarding
A few days with your team to learn people, systems and domain.
- 03
Remote delivery
Mostly remote, with fixed rituals and visits as needed.
- 04
Handover & enablement
Documentation and knowledge transfer – your team takes over.
faq
Frequently asked questions
Do we need our own AI models?
In most cases, no. Modern language models via API, combined with your data through RAG and good prompt and tool design, cover most use cases. Custom models or fine-tuning only pay off for very specific requirements.
How do you make sure the AI doesn't give wrong answers?
By grounding it in your sources with citations, setting clear boundaries in prompt and tool design and, above all, through evaluation: we define test cases and measure quality on every change instead of relying on gut feeling.
What does it cost to run an AI integration?
That depends heavily on volume. I plan for cost from day one – choosing the right model, caching, shorter contexts and batching often reduce running costs significantly.
Can our data stay in the EU?
Yes. Many providers offer EU regions and data processing agreements that exclude training on your data. Where needed, we also evaluate self-hosted models.
How long does a first prototype take?
A focused prototype against real data typically takes two to four weeks – including success metrics on which you can base your decision about the next step.
contact
Let's talk about your project.
Tell me briefly what it's about – I'll get back to you within 1–2 business days.