Nagamese AI Experiments
Exploring local language technology.
Tools that help AI write Nagamese, the common language of Nagaland, where the same word is often spelled several ways.
- When
- Since July 2025
- Status
- Research prototype
- Sheet
Overview
Nagamese is the common language of Nagaland, and the same word is often spelled several ways. In my testing, general AI models handle it poorly, so I'm building tools that give the model what it's missing.
I went to school in Dimapur, and I wanted to see how far today's tools can go.
What I built
Generator frontend
- A React 19 and Vite app with Tailwind CSS where a business writes a post in English and gets Nagamese copy for Facebook, Instagram, WhatsApp, email or print.
Generator backend
- A Node.js and Express API on MongoDB, with JWT auth and a Redis cache.
- An English–Nagamese dictionary of words and phrases, with an admin API to edit it, and a cache of past translations.
AI pipeline
- Gemini 2.0 Flash, with OpenAI as fallback.
- A translate path that first simplifies the English, then adds matching dictionary entries to the prompt as a glossary.
Translator, with Sahil
- An English, Hindi and Nagamese translation API in FastAPI. We designed the approach, the training data, the prompts and the API together, and Sahil wrote the code. It has a GPT path through LangChain and a model of our own in progress.
How it works
- English
- Simplify15 words or fewer
- Glossaryfrom my dictionary
- Gemini
- Nagamese
- English
- GeminiOpenAI as fallback
- Nagamese
On the translate path, English is simplified into short sentences, matching dictionary entries are added to the prompt as a glossary, and Gemini writes the Nagamese. Post generation sends the English straight to Gemini, with OpenAI as fallback.
Key decisions
Simplify before translating
On the translate path, the English is first rewritten into short, active sentences of 15 words or fewer, so the model works with plain sentences.
In Fig. 1:Simplify
Hand the model a glossary
For each translation, matching entries from my English–Nagamese dictionary go into the Gemini prompt, so the output uses the dictionary's spellings instead of guessing.
In Fig. 1:Glossary
What's next
A review set judged by native speakers, because a single score can't capture spelling that varies from writer to writer. Results go up here once that review exists.
Stack
Frontend
Backend
AI
Translator