Meet Lectorium: Exploring AI That Speaks Estonian - and Reads the Law

15 April, 2026

Every so often, a project starts not from a customer brief but from a question that refuses to go away. Ours was simple: what would it take to build a small, purposeful AI - trained from scratch, by us - that actually understands Estonian?

That question has quietly turned into something we are calling Lectorium.

Why Estonian, and why now

Most of the AI models making headlines today are built by enormous teams, on enormous budgets, for a global audience. English gets the lion's share of attention. Estonian, a small and beautifully complex language, tends to sit at the margins of these systems — supported, but rarely a first-class citizen.

We think that is a missed opportunity. Estonia has a rich, well-digitised public information landscape, a culture that embraces technology, and a language that deserves tools built with it in mind, not around it. At Crowned Phoenix, we wanted to find out what happens when a small team treats Estonian not as an afterthought, but as the starting point.

What Lectorium is (and isn't)

Lectorium is our exploration into domain-specific AI for Estonia. The working concept is a reading companion for Estonian law — a system that can take a plain-language question about legislation and return a grounded, honest answer backed by references to the actual legal text in Riigi Teataja.

The emphasis is on the word grounded. We are not trying to build a model that memorises every Estonian statute and recites it from memory — that path leads to confident-sounding but unreliable answers, which is the last thing anyone wants when the subject is the law. Instead, Lectorium is being designed to retrieve the relevant articles at the moment of asking and to cite them clearly in its response, so that any user can click through and read the source themselves.

Transparency is the whole point. An answer without a citation is, to us, not an answer worth giving.

What we are actually investigating

This project is as much a research journey as a product effort. Along the way we are exploring a set of questions that, honestly, we could not fully answer before we started:

• How small can a language model be and still produce coherent Estonian?
• How much does a tokenizer purpose-built for Estonian morphology matter in practice?
• What does it take to turn a publicly available legal corpus into something a model can reason over reliably?
• How do you combine retrieval with generation so the system is accurate and useful, not just one or the other?
• What is the right way to measure success when the domain is legal — where being almost right is often worse than being silent?

These are not rhetorical questions. We will be testing, measuring, and adjusting as we go.

A deliberately small model

Much of the current conversation about AI is dominated by scale — bigger models, more GPUs, larger training runs. Lectorium is a deliberate step in the other direction. We are building a small language model: something trainable on modest hardware, possible to inspect and understand end to end, and cheap enough to run that it does not require a fleet of cloud accelerators to serve a single question.

Small, in our view, is not a limitation. It is a design choice. A focused model that does one thing well, on a language and a domain it was built for, is a different animal than a general-purpose giant trying to be everything to everyone.

The road ahead

We are at the early stages. The first milestones are foundational: training a general-purpose Estonian model that produces readable text, building the ingest pipeline that pulls legislation from Riigi Teataja into a searchable form, and laying down the evaluation framework we will use to keep ourselves honest about what works.

From there, the pieces come together — retrieval, generation, citations, a usable interface — and Lectorium starts to look less like an experiment and more like a tool.

We will be sharing what we learn as we go: what surprised us, what failed, what turned out to be harder or easier than expected. If you are curious about language models, Estonian-first technology, or the messy middle ground between AI research and real-world products, we would love to have you follow along.

More soon, from the Crowned Phoenix workshop.