Text to Speech Model (Dhivehi)
Javaabu
An end-to-end model that can take dhivehi letters such as ހ، ށ، ނ directly and tokenize them at character level and then synthesize speech.
It is currenlty deployed as a service on www.dhivehigpt.com/tts. A smaller model of around 10 MB
of this model is deployed on a mobile version that's currently being tested by the blind and visually impaired community of Maldives. And in the process
made an open source contribution to Sherpa-onnx. Quoted
at https://github.com/k2-fsa/sherpa-onnx/issues/1088. And the original contribution https://github.com/k2-fsa/sherpa-onnx/discussions/890
Authored the mobile app as well, It is written in kotlin.
OCR Model (Dhivehi)
Javaabu
A model that can recognize dhivehi text or handwriting with incredible accuracy or character error rate. A model that is currently the best in the world at dhivehi recognition (even better google vision). Deployment currently in progress and will be available on dhivehigpt soon.
Working to create a Dhivehi LLM
Javaabu
Finetuned the famous Generatively Pretrained Transformer (GPT-2) open sourced by Open AI by using dhivehi data available on online Maldivian news websites.
AI HS: Harmonized System Code Recommendation
Maldives Customs Service
AI HS is an advanced model trained on over 2.5 million historical records from the Maldives Customs Service (MCS), designed to recommend Harmonized System (HS) Codes based on commercial product descriptions. Explore the tool at www.customs.gov.mv/eServices/findtariff.
Automated Document Management Bot
Maldives Customs Service
A sophisticated bot facilitating the upload and download of supporting documents related to customs declarations from the ASYCUDA database. It efficiently handles various documents including invoices, packing lists, and bills of lading.