Project NoCap

Social media are flooded with false information and half-truths. NoCap is an Instagram bot; you send us a post or a reel, we cap-check it with online sources and return a comprehensive analysis.

Comment

Inspiration

As a part of our spring college class, our team of five friends faced the ever-present challenge: "How might we fight against the tumultuous amount of misinformation online?" We spent considerable time researching the topic. It quickly became clear just how tricky the issue is: it's not enough merely to detect fake news and disprove it, but people also need to care enough and be motivated to fact-check.

Find out more about us on our web!

What it does

After conducting research and multiple user interviews, we realized that our solution needs to be super easy to use. We chose to focus on Instagram, one of the major social media platforms where many young people consume information and daily news. Thus we created NoCap.

NoCap is an AI-powered fact-checking assistant integrated directly into Instagram. Users can forward any post or reel to our chatbot @project_nocap) and receive a detailed verification report within two minutes. The report includes:

  • Real-time Analysis: Instant fact-checking results powered by advanced AI technology.
  • Bias Detection: Identification of potential biases and hidden agendas in content.
  • Additional Sources: Curated list of reliable sources and references for further reading, categorized based on their reliability as per established fact-checkers.
  • Multilingual support: Wide variety of languages supported

This seamless integration allows users to verify information without leaving the Instagram platform, promoting informed decision-making in a user-friendly manner.

How we built it

Our tool is in its foundations an API, that integrates with Instagram and handles any message we receive through Instagram API Webhooks.

AI models play a crucial role throughout the whole process. We have built our own "crew" that handles each message in the following steps:

  1. Information extraction: we use AI models to describe the content of a post image/ transcribe reels and describe its visual content
  2. Claims analysis: Perplexity Sonar is crucial for our project. With its advanced search capabilities, we are able to analyze the extracted information, search for it online, and find relevant sources to support the analysis.
  3. Report writing: in the last step, we summarize our analysis in a concise and clear way, which we send back to the user again through IG API

Challenges we ran into

There are 3 main considerations we have when designing the responses: precision, speed and cost. We need to balance the three which has proven quite challenging.

As with every project, there were many technical challenges:

  • Instagram API limitations: there is only a specific information that Instagram shares with its users, which can be quite limiting
  • Fact-checking: Using publicly available AI models limits control over outputs. Given the diverse content types we receive, from memes to political news to hate speech, we continuously refine our prompts to maintain accuracy and credibility
  • Hallucinations: Grounding our fact-checking process in Perplexity Sonar search was very helpful as we actually got functioning, relevant, real sources to support the analysis
  • Speed: Since our project utilizes multiple agents through API, we are trying to optimize and think of ways to make the process faster. Right now we take about 60s on average to respond, however we would like to cut down the time significantly. It is not necessarily a priority tho, as users can share the post with us, continue scrolling on their feed and they get notified of an incoming message.
  • Computing capacity: Since we are very low-cost, we do not have much computing power available, we have to design our code with efficiency in mind.
  • Reliability: Originally we thought of using already existing multi-agent libraries to streamline our workflow. That however caused many reliability issues as we got frequent errors. Thus we switched to a very basic agent workflow.

Accomplishments that we're proud of

!! IT ACTAULLY WORKS !!

We got to the point where we are reliably able to respond to the majority of the queries we receive on our instagram profile. The answers are often helpful and relevant (although we are still very much fine-tuning them).

What we learned

Almost everything. We have never used FastAPI before, we had almost no experience with building AI agent workflows, we never used IG API, and so on... It has been a great learning journey that has given us so much.

What's next for Project NoCap

There are multiple pathways our project can go in:

  • Our main priority is to optimize the fact-checking process, we want better, more reliable and relevant responses to each post!
  • Integrating new models. The AI world is moving fast, we want to stay up-to-date and use the new technology to improve the project
  • Expanding to new platforms. Sadly, fake news do not live only on Instagram, there are many platforms our project could expand to!
  • Community engagement is one of our main goal into the future. We need to gain recognition and gain a solid user base.

Built With

Share this project:

Updates