Each task has its own Akuru model. They are trained on Dhivehi data, scored on fixed Dhivehi test sets, and replaced only when a new version beats the old one.
akuru-raalhu-P-1.0akuru-bathheli-T-1.0akuru-loa-T-2.0akuru-dhoni-T-1.1akuru-beruakuru-odiSpeech, translation, OCR and chat each have a model trained for that job. Small, fast models that can run on a single GPU and scale to zero.
Dhoni translates a question into English with Bathheli, reasons with an open model, and translates the answer back. It works today while our own Dhivehi LLM is trained.
Numbers, dates, times, money, atoll codes and abbreviations are rewritten the way Maldivians say them, so speech sounds natural.
Testers like, correct and record answers in the app. Approved corrections and recordings go into the next training run.
Models run on GPU workers that start on demand and scale to zero, so idle time costs nothing.
Training runs on A100 80 GB GPUs, about 9× faster than our workstation, started per job and stopped after.
The app, accounts, payments and feedback run on Cloudflare's network, close to users in the Maldives.