mukosozi/mukosozi

Open-source Kinyarwanda grammar correction tool: rule-based text correction for the language of ~12M speakers. # Mukosozi > *Open-source Kinyarwanda grammar correction tool — rule-based text correction for the language of ~12M speakers.* **Mukosozi** (Kinyarwanda: *the corrector, the one who corrects*) is a Kinyarwanda text correction tool. It targets punctuation, spacing, capitalization, and morphological errors that affect written Kinyarwanda — a Bantu language spoken by approximately 12 million people, with limited NLP tooling available. The project is **authored and maintained by a native Kinyarwanda speaker** and grounded in explicit, documented linguistic rules — not opaque models. Every correction is traceable to a rule with a clear, reviewable explanation. --- ## What it does (v0.1.0) Current capabilities, all rule-based and running client-side in the browser: - **Punctuation correction** — spacing around `.`, `,`, `!`, `?`, removal of duplicates - **Capitalization correction** — sentence-initial caps, proper nouns (place names, days, months, common religious and geographic terms) - **Spacing correction** — common merge/split errors (e.g. `nukuri` → `ni ukuri`) - **Lexical exception handling** — words with special rules (e.g. `ubukwe`, which has no plural; `Kabgayi`, which has a unique historical spelling) - **Quotative marker agreement** — subject-prefix concord (`baravuga bati`, not `baravuga ati`) --- ## Quick start Mukosozi v0.1.0 is a single-file web application. No build step, no dependencies. ```bash git clone github.com cd mukosozi open index.html # macOS — or just double-click the file ``` Type or paste Kinyarwanda text on the left; corrected text appears on the right. --- ## Screenshots ### Default state The Mukosozi interface on first load. Bilingual labels in Kinyarwanda and English, dark theme with gold accents inspired by Imigongo geometric tradition, and a clear legend mapping correction types to colors. ### Single correction example A one-word input — `nukuri` — triggers rule **SP-001**: th …