The question
Could a machine write football coverage worth publishing?
Emil had followed the language models since they appeared, and he kept returning to one question. Could they write match previews that someone would publish under their own name, on a deadline, for readers who know the game better than the writer does? He put the question to Simon Lindell at Sportredaktionen in Gothenburg, known to Swedish football as “Dybban”, and got a yes to trying.
The craft
Making a model write like someone who watched the match.
A model writes a football article in seconds, and everyone can tell. Emil had spent years on the other side of that problem, working out how to prompt the models into narrative writing with real language, the kind a reader follows to the end. No single instruction does it. It is technique built up in layers, tested on readers and kept only when it holds. That work became the voices in the system, and it is still where most of his hours go.
The wall
The prose got there. The facts did not.
However good the prompts became, two things kept breaking. The writer would state something the match data did not support, and it would invent words, most often in Swedish, where the models have far less to lean on. Nobody can publish a text that is beautifully written and wrong. So the problem moved from how to write to how to make every claim traceable to a source.
The rebuilds
Platforms, chains, rewrites.
We built the production chain on one platform after another, and rebuilt it every time it hit a ceiling. The steps needed data the platform could not fetch, checks it could not run, or control over the model calls it did not give us. Each rebuild taught us what the software had to be able to do.
Our own platform
Full control over every step.
So we built the platform ourselves. Every step in the chain is our own code: what data comes in, how it is curated, which model runs where, and what has to be true before a text may leave. Since then it has held. The system is in daily production and the articles go out as delivered. The newest thing it produces is video. A customer's own verdict →
Who are we?
We are Emil and Pernilla Ekström, and we build Sportsjournalist together. Emil has worked with AI models since the very beginning, and the voices in the system are his craft. He has a background as a professional poker player, won over a million dollars and went on to coach entrepreneurs and other poker players. Pernilla comes from product management and has worked with several successful tech startups. She built the software that runs the engine and the user portal.
