Guardian

Inspiration

We are Kasia (UX Designer and Researcher) and Javier (Product Manager) working in the area of Artificial Intelligence and with great passion for it. Nevertheless, we also observe some people being afraid of the Internet, the scams and the new wave of revolutionary technology enable through GPT-like models and similar architectures. We have decided to focus on:

Accessibility, by enabling users to access the power of an LLM through already known applications such as Whatsapp. ## What it does The prototype we have build is pretty simple. You can send a URL to a Whatsapp Bot and the LLM will read your Terms of Service, summarize, and give you some red and orange flags. This idea is not new, and heavily inspired by Terms of Service; Didn't Read (ToS;DR). ## How we built it We are using Haystack as the framework, but run out of time to built something more complex with a RAG system or some agentic behaviour, which are our next step. ## Challenges we ran into We started with the idea of privacy-first application, running with Ollama in the browser with Gemma:2b. Unfortunately, we realized that the results were not as we expected so we decided to move to gemini-pro model. As mentioned, the Jupiter Notebook is pretty simple and does not have anything special other than some time of Prompt Engineering. ## Accomplishments that we're proud of We are proud of having done a user research for guiding our decision making. That informed us not to build a new interface, but to re-use the things that the users are already used to. We are also proud of our ongoing research to understand better which users feel more insecure and to which "attacks" they are more vulnerable, and how can we help them with a Guardian (the good AI) beside them. ## What we learned We learned how easy is to use the vertexai api and how plug-and-play is the haystack framework. ## What's next for Guardian Difficult to say, we are still researching and asking for Feedback for our experience. But definitely phishing, attacks through SMS, and similar things, as these kind of attacks are where both less-vulnerable and more-vulnerable user groups shared that are being attacked, and also we believe that AI is going to "augment" and "facilitate" this kind of attacks.