I run M21Global, a translation and localisation company I co-founded in 2005 and have helped run for most of the time since. I am not an engineer by training. My degree is in public relations and media studies, from the University of Exeter, and for most of my career my job was the business of translation: clients, quality, the team, the deadlines.
That changed when generative AI started taking the work. Rather than wait to see how far it would go, I taught myself to build with it. Over the past two years I have rebuilt M21’s production engine around large language models, on my own, with code: a multi-model translation pipeline with a consensus loop, automated quality estimation running on a GPU of my own, deterministic safeguards, and the grading system that decides which segments a human still needs to see. It runs in production, on real client work, measured against what our reviewers actually change.
This site is where I write about that, in the open. The real method, the real code, the failures included. I write it first here and syndicate the business half to LinkedIn. The audience I have in mind is the one I belong to: people running language companies through this shift, and people building production systems with LLMs who want more than a demo.
I am based in Lisbon. I run, currently training toward a marathon, which is the closest thing I have to a method for hard problems: keep going when it stops being comfortable. I speak Portuguese and English fluently, French and Spanish enough to get by.
If any of this is useful to you, or you want to compare notes, I am easiest to reach on LinkedIn or by email at me@diogoheleno.com.