allows users to generate confessions similar to those of a popular Adriatic region website Users can provide an optional prefix and post these confessions under a pseudonym. Also, all users can react to generated confessions.’s frontend is a react-based single-page app. The design is mobile first and uses bootstrap for the main UI components.

The backend service is written in flask and backed by SQLite. The task of confession generation was extracted and assigned its own process due to the large demands on cpu and memory. Inter-process communication was used in the form of message queues handled by rabbitmq.

The generation process was built using a gpt-2 library gpt-2-simple. With its help the 355M hyperparameter gpt-2 model was tuned to our data, which was taken directly from the source website. A total of around 120K confessions were used.

Currently the website backend is not running due to its high running costs.