Language technology for the languages large models forget — Breton, Welsh, and everything else without a corpus.
Applied AI that reaches production: retrieval, LLM workflows, evaluation, and the governance around them.
Software built to ship — SaaS platforms, APIs, data models, deployment.
Teaching and instructional design, from a training day to an adaptive learning system.
What those four senses mean in practice.
Two things I built and still maintain.
Course platform for language schools fluens.co
After Eienn.bzh, my first attempt at a Breton learning platform, Fluens generalised it: a CMS where schools and publishers author their own courses, a learner-facing app, analytics, and Stripe subscriptions. Built end to end — Vue 3 with Radix, Firebase, Node functions.
Crowdsourced omnilingual dictionary leksis.eu source
Leksis is a dictionary for every language, built on the premise that low-resource languages are better served by densely annotated data than by more unstructured text. Contributors publish entries, senses, translations, cognates, sources and inflection paradigms as signed records on their own AT Protocol repository; an AppView indexes them into a graph that can be traversed for indirect translation and exported as spellchecker dictionaries, translation memory or training data. Nobody owns an entry — a record proves authorship, not ownership, and every version is kept.
Open experiments on languages that do not have enough data. Several of them failed; the failures are written up too.
A small retrieval pipeline trained on my own writing. Ask it about me and see how far it gets.
My fork of the open-source grammar checker, adding Welsh rules and dictionaries.
Fine-tuning and evaluating a translation model on a parallel corpus I built by scanning novels.
I work where linguistics, pedagogy and software meet. Most of my time goes to languages computers were never built for — Breton, Welsh — because the problems there are harder, and the tools that come out of them turn out to be better. The rest goes to helping organisations use AI without buying something they cannot maintain.
Consulting, collaboration, or a talk — email is fastest.
