The problem Intercept AI solves

In recent times, we have seen a drastic increase in online toxic content, hate speeches targeting a community,misleading tweets ,and non-contextual and abusive comments, which could be a great start to fire up Protests and Riots. CAA-NRC and Farmer's Bill are great examples of such cases.

Also, We have observed a great increase in violent and pornography images, making the place inappropriate for users.

As a Solution, We came up with our Deep Learning Solution, INTERCEPT AI, a B2B model acting as a layer between user interface and backend. That can filter all kinds of Inappropriate and Toxic Contents and warns the user regarding the act.

Features that we offer:

• Specific Word Restriction: Using OpenAI GPT-3 Embeddings tool to compare input text context similarity with Blocked Words

• Multilingual Abuse Detection: Trained on MURIL BERT model and is able to detect abusive content of different Indian languages like Hindi, Marathi, Assamese, Kanada, Malayalam and many more.

• Violent Adult Content Detection: Trained using CNN ,Transfer learnIng using RESNET50

• Toxic English Hate Speech Detection: Trained using Bi-Directional LSTMs

Challenges we ran into

• Training the model with such huge data was a very challenging task on our system. So we used Kaggle and Google Colab GPUs.

• Deployment of a MURIL BERT model was a bit challenging task.

• Deployment of English Toxic Hate Speech Model faced the difficulty of determining Tokenizer while deploying, So we passed and used Saved Tokenizer that we used for training.