Akuruއަކުރު AI
ދިވެހިOpen the app

The master plan

Make Dhivehi a first-class language for AI: everything a person or an organisation can do with AI in English, they should be able to do in Dhivehi, on models that we own and improve.

Five phases

  1. Foundation

    Mid – October 2026Done

    Our own model for every core Dhivehi task, in a live app.

    • Speech to text, translation, Thaana OCR and chat models trained and serving
    • Dhivehi-first web app with accounts, free monthly credits and payments
    • Cloud training and serverless serving

  2. Quality

    November 2026Now

    Each model good enough to rely on every day.

    • Translator v3: chrF ≥ 70 English → Dhivehi, ≥ 63 Dhivehi → English
    • Speech to text: character error under 15% (Whisper large-v3 full fine-tune)
    • Chat answers grounded in trusted Dhivehi sources instead of guesses
    • A natural Akuru voice for text to speech
    • Weekly native-speaker reviews of every model

  3. A Dhivehi language model

    Nov – Dec 2026Next

    A model that reads, reasons and writes in Dhivehi directly.

    • Continued pre-training of an 8–12B open model in Dhivehi
    • Instruction tuning on reviewed Dhivehi examples and tester feedback
    • Ships as Dhoni v2 only if it beats the pipeline and frontier models in native ratings
    • Then Odi: a larger Dhivehi model

  4. Platform

    2027Later

    Akuru as the Dhivehi AI layer other products build on.

    • API for chat, speech, translation and OCR
    • Document digitisation for courts, councils, ministries and archives
    • Speech for call centres, media captioning and accessibility
    • Private deployments for organisations that need data to stay in the country

  5. Archive and voice

    2027 and beyondLater

    Bring Dhivehi's written and spoken heritage into the digital age.

    • Digitise printed Dhivehi books and records with Loa
    • Studio-recorded Akuru voices with licensed voice actors

The flywheel

1

People use Akuru in Dhivehi

2

Corrections and recordings come back (with consent)

3

Models improve on fixed tests

4

Better models bring more people and organisations

Every round of feedback makes the next version better.

Principles

Dhivehi first

Thaana comes first in every product, English alongside.

Own the models

We train and serve our own models so quality, cost and data stay in our hands.

Measure honestly

Fixed test sets, before-and-after scores, and nothing ships unless it is better.

Respect the data

Consent for recordings, personal data scrubbed from training text, licences tracked.