I ran fivethings at once.Now I run one.

PayIvy was a marketplace for selling digital goods, and it took Bitcoin and altcoins in 2014, when that was still a strange thing to build. For a year it was the only thing I did. Ten thousand sellers signed up. Two million dollars moved through it in six months. In July 2015, NewNote Financial bought it.

I was fast, not good. The failure modes of that system were other people’s money, in a currency nobody could reverse. I have never since been able to treat a payment path as an implementation detail.

The obvious move after an exit is another company. I enrolled in a computer science degree instead. Three years of the things I had been getting away without: complexity, data structures, why a database gives the guarantees it gives.

CodeChamp ran alongside it for six years. Small applications and websites for local businesses, one client at a time. It paid for the degree, and it kept the loop short. Nobody explains a bad abstraction faster than a shopkeeper who cannot find the price field.

SmartWithFood had a document database being asked a question it was never shaped for. Compare two products by ingredient. Then by nutrition. Then search across both. With a senior architect I moved it onto a graph model in Neo4j and rebuilt the schema around the queries people actually ran.

The part I am still proud of is smaller. I built versioned Docker images of the database itself, so a migration that produced bad data could be rolled back to a known state instead of argued about. It was the first time I designed for being wrong.

TXCare builds and maintains MPLS networks. It is over eight million lines of code. Nobody holds a system that size in their head. You work by evidence instead of intuition. You change one thing. You accept that the codebase knows more about the domain than you do.

In the same stretch, nine colleagues and I built Greenr in our own time — an app to make the company argue about its own environmental impact. I led it technically and as a team. Two very different lessons about scale, running in parallel.

This is the peak, and it was not a plan. The integration layer at Axxes was the real work. Every application publishes domain events. Every application gets its own queue. Each one declares exactly which event types it may send and receive. I shipped it as NuGet packages, so a team could join the bus in an afternoon instead of a sprint.

Around it: the Axxes mobile app and Home portal in React Native, React and .NET, including the store releases, the Figma, and the conversations with the people who would use it. And Terraform, Cloudflare and Azure at AdminPulse. Four roles for two and a half years, five at the peak. I would not do it again. I would not trade what it taught me about what only I could do.

A general model pointed at a specific business always feels almost right. The value is not the model. It is the context — the company’s own data acting as its memory, so the system gets less generic every month instead of more.

So the practice is deliberately unglamorous. Stay vendor-agnostic; the best model in eighteen months is not the best model today. Integrate with the data that already exists. Run the hosting, so the client spends attention on the business case and not on inference. Start with small use cases that visibly work.

Jasper D’Hooghe and Sam Van Praet founded Nexa Legal in Antwerp in August 2025 to take the repetitive work out of litigation. I built the AI architecture, came in as a shareholder and became CTO. In February 2026 we raised a million euros from RDY Ventures, Volta Ventures and PMV.

What the product does is narrow on purpose. The Case Flow Wizard walks a litigator through a case: read the intake, build the timeline from three hundred documents, research it against the databases and the firm’s own archive, argue the other side back at you. Research runs about five times faster. The same number of lawyers gets through roughly twice the files.

The engineering that matters is the part nobody demos. Every answer links to its source. A second model sits in judgment on the first. Pseudonymisation happens locally and processing stays in the EU. We switched the automatic strategy generation off and replaced it with a guided process, because the lawyer has to stay liable for the argument, and that means the lawyer has to stay in it.