AI and LLM integration

LLMs are easy to demo and surprisingly hard to ship. I help teams cross that gap: pick the right model, build evaluation harnesses, and design around cost, latency, and the ways these systems fail.

Based in Groningen, working across the Netherlands and remotely for clients worldwide.

LLM-powered features for existing apps and new products: agent workflows, RAG pipelines, and integrations with OpenAI, Anthropic, and similar providers.

Integrations that make it to production. Prompt engineering, evaluation harnesses, and the unglamorous parts of running LLMs in production: cost, latency, and failure modes.

Good for

How I approach it

  1. Start from the workflow. The question is never which model is newest, but where language models actually remove work in your process. I look for the steps where they help and am honest about where they will not.
  2. Prototype on your real data. A small working prototype against your actual documents or tickets says more than any slide deck. That is usually the first milestone.
  3. Evaluate before shipping. LLM features fail in quiet ways, so I build evaluation sets and regression checks. A prompt or model change that makes things worse gets caught before your users catch it.
  4. Plan for the failure cases. Fallbacks, cost and latency budgets, and human review where the stakes require it. The goal is a feature you can put in front of customers.

Work I can show

Tech I work with

Anthropic Claude, OpenAI, embeddings, vector databases, evaluation frameworks, prompt caching.

Frequently asked questions

Will my data be used to train a model?

Not by default. Under Anthropic's and OpenAI's current API terms, your input is not used for training. For private workloads we can also self-host an open-source model so the data never leaves your infrastructure.

How do you measure whether the AI feature is good enough?

With an evaluation harness built from real examples, ideally collected from your domain. Each release is scored against held-out cases for accuracy, latency, and cost so regressions show up before users do.

What about hallucinations and reliability?

I treat them as a design constraint. Outputs are validated, fallbacks are explicit, and high-stakes flows have a human review step. The goal is a feature that fails predictably rather than one that fails invisibly.

Can my data stay private or in the EU?

Yes. Depending on your requirements I use EU-hosted endpoints, self-hosted open-source models, or design the feature so sensitive data never leaves your systems. Privacy constraints go into the design from the start, which is also how GDPR requirements are usually met in practice.

When is an LLM the wrong tool?

When the task needs guaranteed exact results, when a simple rule or query already solves it, or when the volume makes per-request costs unreasonable. If that is your case I will say so and suggest something simpler.

Can you add AI to an existing app?

Yes, and it is a lot of what I do. Often the right move is one focused feature added to software that already exists, a search box, a summarizer, a support assistant, rather than rebuilding the product around a model.

Want to talk it through?

Tell me what you're trying to do. I'll let you know honestly whether I'm a good fit.

Get in touch

Other services

Start a conversation