Safebook

About Page
Landing Page
Post History
Post Page

Inspiration: As we witness the rapid digitization of our world, it's become increasingly apparent that seniors often find themselves at a disadvantage when it comes to navigating the complexities of technology. As a result, this demographic is one of the most vulnerable to online scams. To address this issue, we were inspired to create an intuitive and simple application that would help seniors identify potential scams, allowing them to connect to their loved ones on the internet more safely.

What it does: Safebook provides an easy-to-use interface allowing you to input any text that needs to be validated. It uses a post interface to simulate verifying social media posts. It analyzes the message using an AI model we developed and informs you of the likelihood of that message being a scam.

How we built it: Using PyTorch, we based our AI off of Google's powerful BERT transformer equipped with powerful text analysis capabilities. We added a few additional classifier layers and trained it on a rich dataset of over 100,000 example scam messages. We kept the learning rate high for our added layers, while making it very low for Google's pre-trained layers, a process known as fine-tuning. After training, we achieved an impressive 99% accuracy at detecting scam messages from regular ones in our test dataset. We hooked our AI up with an efficient interface using FastAPI, allowing it to connect with our frontend. The frontend is built with React.js and provides a simple and intuitive interface to the model for checking potential scam messages.

Challenges we ran into: We wanted to make the AI as powerful as possible, so balancing our limited computing resources with maximizing the capabilities of the model was an important challenge to overcome. In addition, we had several extra ideas, such as making a chrome extension that could integrate with Gmail to provide a more automated experience, but we were unfortunately not able to get it ready in time.

Accomplishments that we're proud of: We're very happy with the high level of accuracy we were able to obtain with the AI as well as the seamless experience our interface provides to seniors checking for scams. In addition, the performance of the server-side API is excellent, and we're proud of how smoothly the stack fits together.

What we learned: We learned a lot about large language models and text analysis. In addition, we were able to bolster our web development experience using FastAPI, React, and Tailwind.

What's next for winhacks2024: If we were to continue with the project, we would create additional interfaces to the model such as making a Chrome extension that could integrate with Gmail to provide a more automated experience. The model is also trained on e-mails, making it less accurate with shorter text. To improve the project, we would retrain it including shorter pieces of text, like personal messages.