The problem Samachar AI solves

A Fake News Detection App is a need now a days because there's a lot of false information out there. It helps us figure out what's true and what's not. Sometimes, news stories try to trick us or spread hate, defamation and the misinformation and the app can catch those too. By using it, we can make sure we're getting the real facts and not being misled by fake news. It's like having a tool to separate truth from lies in the news we see online.

We team - Epoch Builders came up with the solution of the Fake News Detection Application by proposing a new architecture of telling whether the news is fake or not. There is no direct way to say that, so we built multiple models like Hate Speech Detection and Propaganda Detection to understand the instrinsic inside the text of the Article. The feature comprises of the summarization, so that you don't need to go through the article to evlaute. As the Bert model converts it into the pointers and then provides the sentiment analysis of the article text.

Now, Comes the part of understanding whether the news provided had an unhealthy impact towards the society by having the hatespeech or the offensive language. At the end, checking the propaganda Detection which is been trained on the QCRI Propaganda Dataset and all these models are deployed on the AWS INTEL ICE LAKE 4GB Server via HuggingFace.

All the Analytics generated is provided to the Gemini Model to make an intution out of it. As of to follow the news article or not. It guides the user. The model hosted on the High RAM serves better speed and the solution can be employed in the real time and built with the motive to solve the real world problem.

Challenges we ran into

There were many problems we ran into like technical and idea wise. The complications of the initial design on such a time constraint led us to scrape many of our implementation ideas. The technical issues regarding training the model in a scenario where the laptops weren't on charging took a toll on gpu thereby slowing us down. The integration of the saved model in a way that supports the training weights and biases but also is served through torch was a hard task to deal. The deployment of the model on aws took us 2 tries because the debit card failed while subscribing. We tried other alternatives for deployment but AWS was the best option. Some python dependencies were crashing because of the version mismatch and we had to uninstall python several times.

At the end, we were able to solve all the problems and sucessfully completed the project which can be launched and used by the user. We have a believe that it would have a positive impact in the society and ready to get into the hands of the user.