Inspiration
Misinformation has become more prevalent in recent years with the rise of generative AI. As a result more people are prone to ingesting false information. The aim of uBlockAI is to limit the amount of false information which is presented, with the intended audience being senior citizens.
What it does
uBlock AI is a chrome extension which takes Instagram posts and extracts the relevant information through OCR and passes that information along to a Backboard AI agent. This agent then identifies whether this post is meant to spread information (either false or true), and then accordingly verifies claims using alternative news sources.
How we built it
The chrome extensions captures different posts from the social media feeds such as Instagram, and afterwards passes it on to the backend along with the caption and metadata. We capture the OCR text from the images and pass all this information to Multi AI agents which utilizes various tools to verify the claim and produces an AI score, misinformation score, the reasoning it took for agent to come to particular conclusion, and evidence supporting that decision. This gets passed again to the frontend, which then flags and removes the post based on user chosen parameters.
Challenges we ran into
. There were no significant challenges; however, there were many small struggles implementing the agentic logic. One such issue was the latency of the agents. We we're able to manipulate the prompt and model selection to mitigate the amount of time the backend took to verify claims, but there is still significant room for improvement.
Accomplishments that we're proud of
The most surprising aspect of the experience was how well the AI agent can verify information when given a good prompt and tools. Given the short time-frame of the Hackathon, we were very surprised with how well this aspect turned out.
What we learned
- Working with Backboard agents,
- Working with the Instagram DOM
- Working with chrome extension development
- OCR fine-tuning
What's next for uBlockAI
Currently, this only works for Instagram photo posts. The next big ticket item would be to utilize TwelveLabs for videos/reels. After that, adding support for multiple platforms.
Built With
- agents
- backboard
- chrome
- fastapi
- gemini
- gpt
- javascript
- python
- tesseract
Log in or sign up for Devpost to join the conversation.