NagaTranslate translates between English and Nagamese, Ao (Chungli) and Sema (Sümi), languages with very little digital text. I built the whole thing: data, backend, website and Android app. It is in production.
How translation works
Large language models do not know these languages well enough to translate them unaided. Instead, each request is grounded in real examples:
- Retrieval: a BM25 retriever with coverage scoring and spelling-variant folding finds the most relevant sentence pairs and grammar notes in curated corpora for the target language.
- Generation: an LLM translates using those examples and per-language grammar rules.
- Guarding: Nagamese output is checked against real published text, and word forms that never occur there (often borrowed mainland forms) are rewritten.
The corpora are built from public-domain and openly licensed sources (historical grammars, dictionaries, Bible translations), cleaned with a reproducible pipeline and checked against real-world usage in news text.
Beyond translation
- Beta languages: Tangkhul, Angami, Rongmei and Phom, each clearly marked as beta.
- Community data: speakers contribute and correct phrases for 22 Naga languages, which feeds future models.
- Learning guides: long-form pages on grammar and everyday phrases.
Stack
- Vite and React 19 single-page app, prerendered to static HTML for speed and search visibility.
- Vercel Edge Functions for the translation and chat APIs, with fail-closed rate limiting on Vercel KV.
- A Flutter Android app that uses the same API.
Related research: NagaNLP.
Try it: www.nagatranslate.in