Uktubly is a web app for practicing writing Arabic letters. Its purpose is for me to gain hands-on experience with building models designed for practical Arabic handwriting recognition tasks and integrating them into a usable product, while also providing a fun tool to help friends and family with language learning.

Why's it called "Uktubly"? That's a sound-alike for "write for me" in Arabic. :)

The codebase is split into two parts:

The /ml folder contains a script for building and training a CNN implemented with the tfjs-node library. It is trained on a dataset provided by Dr Mohamed Loey on Kaggle. It is a relatively simple model designed to be run on a local machine and isn't very resource-intensive. For reference, the training completes within about 10 minutes on a M1 Macbook Air when the number of epochs is set to 20, but that number can be reduced to about half without much effect.

To run the model training script: cd ml && pnpm i && pnpm train

Note that this doesn't automatically include it in the frontend. There's also a convenience script in the root directory, build-model.sh, which builds the model and copies the output into the frontend app's public folder.

The /frontend folder contains the main React app, which contains a minimal tldraw canvas where the user can draw a letter. The app will try to identify it using the pre-trained model generated by the /ml folder's script which runs natively in the browser, so it's very snappy.

To run the frontend: cd frontend && pnpm i && npm dev

This is a work in progress! The current iteration of the model recognises most letters pretty well. At the moment the application only recognises standalone versions of letters.

• Addition of starting, medial and ending letter forms.
• Addition of numbers.