AI features that hold up in production.
I add AI to products people already use: assistants in support, agents that work through your API, and features that draft, search or extract. At Mollie I built AI into the support team's chat, email and voice channels. I build the feature and the checks around it, so it keeps working after the demo.
I reply personally within 2 business days. Rates come up on a call, once the scope is clear.
What I build
Support assistants
AI in the channels your customers already use, answering from your own content and passing to a person when it should.
Agents over your own API
Agents that act through your product's real operations and permissions, over REST or MCP.
Drafting, search and extraction
Features that write a first draft, find the right record or pull structured data out of documents.
Tested like the rest of the code
Tests and evaluations for the AI paths, so a model change or a prompt edit shows up before your users see it.
A good fit
- A product with real users and a job AI could do for them
- A support team answering the same questions across chat, email or voice
- A prototype that works in a demo and now has to be reliable
Not a fit
- A demo to find out whether AI is interesting at all
- Training or researching new models
How it runs
- 01
You write to me
Tell me what you're building through the form. I reply personally within 2 business days.
- 02
We talk it through
A call on the problem, the fit and what it should cost. Rates come up here, once the scope is clear.
- 03
I scope it in writing
What ships first, what it's built on and how it's tested, agreed before the build starts.
- 04
I build it test-first
Each change starts with a failing test, and you see the product running as it grows.
- 05
It's yours to run
The code, the tests and the notes your team needs to keep building on it.
Proof you can check
- Case studyMollie support AIAfter the website revamp I moved to Support and built AI into Mollie's chat, email and voice channels. Under NDA, so scope only.
- Case studyRelayMy own AI harness: models from four providers chat, write, draw, speak and render in one macOS workspace.
- Case studyTaste VaultAgents turn a website into a build prompt, then a coding agent rebuilds the page from it without seeing the original, to test the prompt.
Questions, answered
Which models do you work with?
The ones that fit the job. Relay runs models from four providers side by side, and I choose per task on quality, speed, price and where your data may go.
How do you stop an AI feature from making things up?
By narrowing what it does and checking it. It answers from your data, says when it doesn't know, and every important path has a test or an evaluation I can rerun when the model changes.
Can you add AI to an existing product?
Yes. That's most of this work: a new feature in a codebase that already has users, without rewriting what works.
What did you build at Mollie?
AI for Mollie's support channels across chat, email and voice, after working as the main engineer on the website revamp. The work is under NDA, so I describe the scope and not the numbers.
Tell me what you're building.
One message is enough to start. I read every one myself and reply within 2 business days.
Need a website for a small business? That's my studio, SharpHaw (opens in a new tab).