Garo AI & Achik AI language research

A·we is where we begin.

A long-term language programme begins by taking one living variety seriously.

A language programme, built with care.

A·chik Mande is not a single, uniform way of speaking. Across the Garo Hills and neighbouring regions, related varieties carry distinct histories, sound systems, vocabularies, and patterns of everyday use. Those differences are part of the language landscape; they are not noise to be removed in the name of standardisation.

The term Garo is widely used in English, while A·chik Mande remains an important self-description within the community. Our research begins from that distinction: a language programme must recognise the people and places that give a variety its meaning before it turns speech or text into a technical resource.

A·we is our first programme. It has an established written role in educational, official, and religious contexts, making it a practical place to establish careful methods. That role does not make A·we a replacement for every other variety. It gives us a defined starting point from which to learn what responsible language research requires.

We are treating A·we as living language, not as a fixed archive. The work will attend to the relationship between written convention and speech, to the importance of pronunciation and the glottal stop, and to the way meaning changes across setting, speaker, and generation. A useful system must remain accountable to those realities.

Our first task is to build a disciplined research foundation. Language material must be traceable to its source and conditions of use; evaluation must test whether a system preserves meaning rather than merely producing fluent-looking text; and important decisions must remain open to review. These are operating requirements, not optional safeguards.

This approach also sets a boundary around what we claim. We will not describe early experiments as a finished language system, and we will not use a benchmark as a substitute for speaker judgement. Progress must be demonstrated through appropriate evaluation and review before it is presented as a result.

Ruga is the next planned programme. Its work will begin with its own linguistic and community context, rather than with an assumption that A·we data or methods can simply be transferred. A related language or variety may offer useful questions, but it cannot remove the need for dedicated study.

As the programme expands, we intend to build reusable research infrastructure for additional languages: tools for data stewardship, evaluation, documentation, and controlled iteration. The purpose is not to make every language look the same. It is to reduce repeated technical work while protecting the distinctions that matter.

Reinforcement learning, agent-based evaluation, and bounded recursive self-improvement are future research directions within that infrastructure. Each would be used only where its goals, evidence, and safeguards can be made clear. The long-term aim is a faster and more rigorous way to support language work—without confusing speed with care.