KAeasy: Georgian writing tools for AI agents

KAeasy is SiTech's in-house toolkit of Georgian writing skills for AI agents. It started from one practical problem: cheap models write a lot of Georgian, but rarely the way a trained editor would publish it.

What it solves

Machine translation fails in Georgian in the same few places every time: Russian and English calques, wrong endings on brand names, awkward number and date forms, passive voice piled up, and non-standard quotation marks. KAeasy turns those failures into rules and examples and hands them to the agent while it writes.

How it works

The core is 14 skills: a writing workflow, grammar, natural style, names and numbers, news, social media, interface strings, speech, proofreading, business correspondence, marketing, official documents, support and translation. Each skill carries the rules, the forbidden and preferred forms, and wrong to right examples.

Four deterministic tools sit beside the skills. ka_lint checks the mechanical rules in four modes: article, social, interface and speech. In publish mode it is the gate before anything goes live. ka_fix repairs what a machine can repair: quotes, brand hyphens, transliteration, decimal commas, dashes and spacing. It never touches grammar, which stays a job for a human or a model.

ka_speech converts written Georgian into speech text: numbers as words, times, ranges, money, units, versions and brands in Georgian letters. That step matters for any voice product, because a TTS engine often reads digits incorrectly. ka_learn and ka_mistakes are a shared memory: every human correction returns to every agent, and the most frequent mistakes surface first.

Why it costs nothing to run

KAeasy calls no model, needs no API key and no network. The knowledge lives in skills, the checking lives in deterministic code. Every call is free, and it still works when a model quota is exhausted.

Where it runs today

The sitech.ge blog uses it across eight locales, the tbilisi.today editorial lane runs through it, and the Veasy video engine sends every narration through the Georgian gate before publishing. The project ships with 120 offline evaluation tasks and a terminology base that has already passed 156 kilobytes.

How to install it

Installation is three steps: clone the repository, link it into a profile's plugins directory and enable the plugin. The agent then calls the skills with the ka prefix and checks the text before publishing.

Read next: Veasy: SiTech’s video engine for agents