About.

Three systems in a year

A lighting catalogue generated from store photographs, a shop's fulfilment console built beside its point of sale without ever writing to it, and a public-records pipeline that can rebuild every interpretation from untouched originals.

A complete payments platform, agent-built

Agents wrote effectively all of it under my direction. I built the machinery that made that defensible: shared memory, an enforced rulebook, verification as a gate rather than a stage.

Turluga: fifteen domains, one ontology

A discovery platform covering thousands of businesses across fifteen very different domains, all under one inherited type system. The extraction was pinned to a schema, so the model filled in a form instead of writing fiction. Built by hand, and the last large system where most of the code came from mine.

CARF Curaçao, a shelter rebuilt in a weekend

An old WordPress dog-adoption site became a swipe-based adoption funnel with real filtering and a WhatsApp call to action. Two days, start to finish, against a legacy shelter API.

WattWatt

The whole site from Figma to production, with Google Calendar and Sheets integrations behind it.

The API year: three tools, built to learn the new shape

A detector for logical fallacies in podcast audio, a sentiment engine that read YouTube comments back to creators, and a text-to-playlist generator built on embeddings and a vector store. Retrieval-augmented generation, while the term was still new.

What I was saying at the time

“A new agent that creates code for us on demand, and before you know it, it's going to build projects as well in the future.”

Said on camera, May 2023. Fifty people watched it.

There was no audience to perform that for, which is rather the point. The same year I also said engineers who could deploy their own architectures would be "irreplaceable" by AI. That was cope. Agents reach into architecture and deployment now too. I keep the full graded list, wrong calls included.

Software Development Engineer Intern, Amazon (AWS)

Built a visualiser for DoWhy, Amazon's causal-inference library, and the ECS/EC2 infrastructure under it. Distributed systems at a scale where carelessness is expensive.

Full-stack systems at Daccs

A reusable React front end that cut launch time by 75%, a shared GraphQL and Node API across projects, an SES migration that cut email costs by 90%, CI/CD from scratch, and a full test suite end to end.

GPT-2 in production, for journalists

Sole developer on a tool that ranked, summarised and pulled keywords out of interview data, running in front of real users. It shipped with GPT-2 and GPT-3 doing the summarising, a year before ChatGPT existed. My CV called it a "pre-GPT AI tool", which undersells it. It was a language model in production when almost nobody had one.

StudyPrep, an AI tutor before that was a category

Turned a student's own lecture notes and PDFs into flashcards and quizzes on demand, with the questions generated rather than written.

A conversational system for the Dutch police

Built at TU Delft: a chatbot that helped spot early signs of mental illness in the way people write. Conversational AI, two and a half years before anyone was talking about it.