Big AI does not serve small languages well
General models handle Dhivehi text unevenly, and before Akuru no available model could read Thaana. The people who rely on Dhivehi every day are left with rough translations.
For investors
The AI layer for Dhivehi
Akuru builds and owns the models that let people and organisations use AI in Dhivehi. Built on more than five months of research and model development, four Akuru models already run in a live product. Here is why we think this is worth backing.
Why now
Building a language stack got cheap and fast
Open models and rented GPUs make it possible for a small team to build competitive models for one language. We trained speech, translation and OCR models in months, not years.
Whoever owns the data and the tests wins
Quality in a small language comes from native data, native reviewers and honest evaluation. That is slow to copy, and it compounds with every user correction and digitised page.
Market
Every Dhivehi speaker
A language of around 400,000 people in the Maldives and on Minicoy, with high smartphone use. Consumers want chat, voice and translation that work in their own language.
Institutions that run on Dhivehi
Government, courts, councils, schools, newsrooms and banks produce and process Dhivehi documents and speech every day: digitisation, transcription, translation and search.
A method that travels
The playbook (data pipeline, specialist models, fixed exams, native review) is built to repeat for other under-served languages.
Business model
Consumer app with credits
Live100 free credits a month, then paid credits by bank transfer with automatic receipt checks.
API
PlannedPay-per-use access to chat, speech, translation and OCR for developers and companies.
Digitisation projects
PlannedTurning archives, court records and publications into searchable Dhivehi text.
Private deployments
PlannedModels running inside an organisation for data that must stay in the country.
What makes Akuru hard to copy
Dhivehi data
Cleaned Dhivehi text and speech data for training and testing, growing all the time.
Own models, own tests
Four models trained by us, each with a fixed Dhivehi exam; nothing ships unless it beats the current version.
Native-first product
Thaana-first interface, Dhivehi text normalisation and local payment rails, designed for how Maldivians actually work.
Speed
A live product with four models, improved every day, on serverless costs that drop to zero when idle.
Traction so far
- Live app in testing: chat, voice, translation and image reading, with accounts and payments
- Speech word error 71.6% → 52.3%; translation chrF 50.4 → 63.9 (English → Dhivehi)
- Thaana OCR at 0.31% character error on a 32-page exam, where no model worked before
- Dhoni chat leads the open models we tested on a 200-question Dhivehi exam
- Cloud training and scale-to-zero serving in place
Use of funds
Native data and review
Licensed recordings, studio voices, reviewed instruction data and paid native-speaker evaluation.
Compute
Training a direct Dhivehi language model and keeping every model on a weekly improvement cycle.
Team
Machine learning, Dhivehi linguists and product engineering.
Go to market
Pilots with institutions, the API and partnerships for distribution.
Talk to us
Tell us a little about yourself and we will reply personally, usually within two working days.
- Investor deck and financial model
- Detailed roadmap and model results
- A call with the founding team