Chat
Ask anything and get answers in natural Dhivehi. Akuru Dhoni is our chat model.
AI made for Dhivehi
AI that speaks, reads and writes Dhivehi.
Akuru is a family of AI models, and an app, built for the people who speak Dhivehi. Chat, voice, translation and Thaana text recognition, trained on Dhivehi data and measured on Dhivehi tests.
5+ monthsof Dhivehi research and model development behind four Akuru models in a live app
0.31%character error rate reading Thaana pages and screenshots
52.3%speech word error, down from 71.6% for the base model
+13.5chrF points for English → Dhivehi translation over the base model
Four tools, one account
Everything works in Thaana first, with English alongside.
Voice
Dhivehi speech to text and text to speech, with numbers, dates and money read the way people say them.
Translate
Dhivehi ⇄ English, plus ten more languages, on our own translation model.
Read images
Pull Thaana text out of scans, photos and screenshots. Most AI tools cannot read Thaana at all.
Why Dhivehi needs its own AI
A script of its own
Thaana is used for one language only. General AI tools treat it as an afterthought; every OCR model we tested failed to read it before we trained our own.
Small language, real needs
Courts, councils, newsrooms, schools and families work in Dhivehi every day. They need tools that get the language right, not a rough translation.
Built here, tested here
Our models are trained on Dhivehi text and speech and scored on fixed Dhivehi test sets that are never used for training.
The Akuru model family
Named after Maldivian boats and sea life.
Dhoni
BetaChat
the traditional boat
Bathheli
BetaTranslation
the trading ship between lands
Raalhu
LiveSpeech to text
the wave: sound coming in
Loa
BetaThaana OCR
the eye, and Loamaafaanu
Beru
PlannedText to speech
the Bodu Beru drum: the voice going out
Odi
PlannedLarge Dhivehi LLM
the ocean-going boat
Moving fast, measured honestly
Akuru builds on more than five months of Dhivehi research and model development. Today it runs four of its own models in a live app, with cloud training on A100 GPUs. Every model is scored before and after each change, and only ships when it beats the one before it.
See the progress logBuilding the AI layer for a language
Dhivehi is our first language, and the method is built to repeat. If you invest in language technology, we would like to show you the plan.
For investors