The Flashcard’it Project: oXya’s Participation in the Google AI Hackathon

Last April, the oXya Canada team had the opportunity to take part in the Google hackathon, which this year was dedicated to generative AI, and in particular to Google’s new artificial intelligence model, Gemini. This 2024 Google AI Hackathon edition challenged participants to create innovative applications by using the capabilities of this model.

The Project’s Origin

After deciding to take part in the AI Hackathon, our team gave the matter some serious thought and decided to select an idea that would exploit Gemini’s capabilities. Quickly, inspiration struck: design an application that generates flashcards, serving as a pedagogical tool to facilitate learning and certification preparation in the workplace, using the power of generative AI.

The Flashcard’it project was born, an application that makes learning more engaging, and the creation of tests and training lessons much faster.

Using Gemini in Flashcard’it

Thanks to the integration of Gemini, Flashcard’it can exploit two distinct approaches to flashcard generation. The first method is based on the use of Gemini’s knowledge, independent of corporate data. The second approach involves providing information in PDF form, to enable the models to generate more accurate, contextualized responses. Of course, it’s not that simple; but we’ll come back to this later.

In concrete terms, it’s possible to generate content in the form of flashcards using precise prompts, for example:

“Create a set of 20 questions with 4 possible answers for each, and only one correct answer indicated by a star.”

So, whether it’s for training or certification or you just want to learn more about a topic, the material is presented in a quiz format.

Providing More Context with RAG

As mentioned, questions and answers can be made more precise by providing the AI model with additional information, such as training documents or internal documentation. This is particularly useful when creating training lessons on a company-specific topic.

Possible options include Fine Tuning, which specializes a generative model for specific tasks by adjusting its parameters with targeted data. This offers more accurate responses, but also requires more resources and can introduce biases if the training data is poorly balanced.

Retrieval Augmented Generation (RAG) is an alternative that combines information retrieval and text generation, enabling dynamic adaptation to queries without the need for costly retraining, although its performance depends on the quality and relevance of accessible information sources.

How does it work? Corporate documents are gathered into a corpus, cut into chunks, converted into vectors, then stored in a vector database. When a query to generate Flash Cards is issued, the RAG model transforms it into numerical vectors, identifies relevant documents based on vector similarity, and a large language model (Gemini) combines the query with the retrieved document extracts to produce the final answer.

This method can reduce the hallucination problem of generative language models. By incorporating factual information retrieved from reliable sources, the RAG anchors its answers in verified data, thus reducing the likelihood of generating incorrect or invented information.

The Future of Flashcard’it at oXya

The hackathon provided us with an ideal opportunity to stimulate ideas and explore new approaches and insights concerning the integration of generative AI. The future of Flashcard’it looks bright and promising at oXya, with plans for internal deployment of the application as early as next autumn.

Following the submission of our hackathon project, our team is working on refining the prototype to create a version better adapted to the corporate world and the specific needs of our employees.

Of course, concerns about data security remain paramount. While Flashcard’it is currently developed for non-sensitive data usage, we’ve given some careful consideration to these concerns; the event served as a valuable testing ground for assessing various security aspects.

AI Skills: A Broader Approach

The Google AI Hackathon is part of a more global approach by oXya, which has already experimented with similar technologies as part of its managed services, where generative AI is used to calculate satisfaction scores. The development of the new application thus represents a new step in this direction for oXya, which aims to further develop its expertise in order to better serve and advise its customers.

The event, which took place last April, was above all an excellent opportunity to collaborate, exchange ideas and strengthen ties between our members. Everyone had the opportunity to express their views in order to move the project forward, all in good team spirit. Learn more about the application by watching the video presentation created for the occasion.

By taking part in events like the Google AI Hackathon, we put our oXya teams at the heart of the latest advances in AI, while keeping a close eye on security and confidentiality issues. Would you like to find out more about our innovative working methods? Don’t hesitate to contact us here!