Political discussion on major social media platforms such as Twitter is often flooded with conflicts and polarization. Users sometimes would use adversarial expressions towards political candidates to undermine their legitimacy or intend to discourage them from competing. Thus, identifying whether the interaction is adversarial between a reply and a tweet and whether the content is direct to the political candidate is essential to step towards a methodical and harmonious online environment. We focus on the direction of adversary observed in the tweets from 2018 US general election period, produced well-formatted datasets which contains more than 1.5 million data points covering tweets, user information and candidate information, and developed multiple models combining heuristics and machine learning techniques to predict adversarial direction. Continuing with last semester’s harassment direction model development, we extended our work to embed the model into the backend of a web service - Twitter Dashboard, in order to help registered users automatically filter adversarial content from his/her Twitter account. We built the web client with Flask framework on Google Cloud Platform. On the server side, we modified the models from direction classification to predicting whether to mute a replier, using logistic regression and BERT models. Users also have the freedom to check muted replies and choose to unmute certain repliers. User tests received satisfactory model performance.