BriBooks improves children’s creative writing with generative AI, powered by AWS

Generative artificial intelligence (generative AI) has the potential to play several important roles in education, transforming the way we teach and learn. This blog post looks at how one EdTech startup, BriBooks, is leveraging generative AI to assist young children with creative writing.

Introducing BriBoo

BriBooks is the world’s leading children’s creative writing platform, enabling kids to learn, write, and publish their books online. BriBoo, an intelligent writing assistant, is built on Amazon Web Services (AWS). It leverages the generative AI model to help young authors with contextual sentence suggestions when they have writer’s block.

This is a screenshot showing how BriBoo, BriBooks’ writing assistant, can help young authors with contextual sentence suggestions when they run into writer’s block.

Personalizing BriBoo

Numerous ready-to-use large language models (LLMs), both third-party and open source, play a crucial role in generating recommendations. These foundational models are capable of providing generic suggestions when prompted. However, BriBooks faced a unique challenge: they required predictions tailored specifically to the writing style of each author on their platform. To address this, they collaborated with AWS to fine-tune a base foundation model using the collective data from the platform’s writers, encompassing more than 100,000 sentences. This data was pre-processed to include additional attributes such as the author’s demographics and the genre of the writing in progress. As a result, the refined model became adept at catering to various demographics, leveraging information from the current sentence, the summary of the story up to that point, and more data.

BriBoo is built customizing GPT-J 6 Billion (an open-source LLM) parameters using Amazon SageMaker, which fine-tunes the model using BriBooks data. This helps the model provide focused recommendations based on an author’s age group, genre of book, and style of writing on their platform.

BriBooks runs large-scale events where children come together to write stories and leverage the BriBoo assistant. With the auto-scaling feature of SageMaker inference, BriBooks assists thousands of children simultaneously leveraging the generative AI assistant. The solution provides highly accurate responses pertinent to the type of story and age of the child.

“Our generative AI based writing assistant was trained on one billion words using Amazon SageMaker. It is trained so well that it has made writing fun, easy, and intuitive for kids,” said Rahul Ranjan, chief business officer of BriBooks.

Implementation approach

BriBooks started with a three week proof-of-concept (POC) with the AWS Build Team. To prove the technical feasibility of the solution, BriBooks trained a model on one genre and grade combination (displayed in Figure 1).

Figure 1. This is an illustrative example of sentence recommendations from BriBoo after an input provides writing genre, student grade level, and a first sentence.

The POC helped establish the training and inference hardware requirements, cost, and quality of generated predictions. Then, an AWS Partner jumped in for the full training and implementation of the solution. 

Figure 2 depicts the architecture used for pre-processing the data stored in Amazon Simple Storage Service (Amazon S3), then using SageMaker Model Training, and eventually deploying it as a SageMaker real-time inference endpoint.

Figure 2. This diagram depicts the architecture used for pre-processing the data stored in Amazon S3, which is then used in SageMaker Model Training, and eventually deployed as a SageMaker Real-Time Inference Endpoint.

Security is job zero at Amazon, so the team ensured security was baked in from the initial stages of the implementation. The data used for training the model was stored in Amazon S3 and completely encrypted using a customer encryption key. SageMaker Studio notebooks and SageMaker notebook instances, along with model-building data and model artifacts, as well as output from training jobs, was encrypted by default using the AWS Key Management Service (AWS KMS) for Amazon S3. SageMaker training notebooks, model artifacts and even inference endpoints are also protected by Amazon Identity and Access Management (IAM) policies controlling access to all these artifacts.

Other use cases

Generative AI will help with additional aspects including automation of data entry and processing tasks. This is particularly useful for categorizing books, updating prices, or managing inventory. Generative AI can also power chatbots that respond to user questions. Since BriBooks operates in multiple countries and serves customers who speak different languages, generative AI can be used for language translation on the website or in customer support interactions.

BriBooks is also looking to leverage multi-modal LLMs to augment text suggestions with image generations to bring visual aspects to stories.

Conclusion

Generative AI significantly aids BriBooks in enhancing their virtual assistant by harnessing the power of their proprietary data. This underscores the pivotal role of data as the foundation of generative AI, as the system relies on comprehensive and unique datasets to refine its understanding and response mechanisms. The proprietary data not only enables the virtual assistant to align more closely with BriBooks’ specific requirements but also contributes to the system’s adaptability and effectiveness. In essence, the success of generative AI applications, such as BriBooks’ virtual assistant, hinges on the richness and relevance of the data they are built upon, emphasizing the paramount importance of high-quality datasets in shaping the capabilities of these advanced AI systems.

Learn more by exploring generative AI on AWS.