Be Right Back

Talk to your friends when they're gone

Comment

Check it out on GitHub!

The machine learning and web app segments are split into 2 different branches. Make sure to switch to these branches to see the source code! You can view the repository here.

Inspiration

Inspired in part by the Black Mirror episode of the same title (though we had similar thoughts before we made the connection).

What it does

The goal of the project is to be able to talk to a neural net simulation of your Facebook friends you've had conversations with. It uses a standard base model and customizes it based on message upload input. However, we ran into some struggles that prevented the full achievement of this goal.

The user downloads their message history data and uploads it to the site. Then, they can theoretically ask the bot to emulate one of their friends and the bot customizes the neural net model to fit the friend in question.

How we built it

Tensor Flow for the machine learning aspect, Node JS and HTML5 for the data-managing website, Python for data scraping. Users can interact with the data through a Facebook Messenger Chat Bot.

Challenges we ran into

AWS wouldn't let us rent a GPU-based E2 instance, and Azure didn't show anything for us either. Thus, training took much longer than expected.

In fact, we had to run back to an apartment at 5 AM to try to run it on a desktop with a GPU... which didn't end up working (as we found out when we got back half an hour after starting the training set).

The Facebook API proved to be more complex than expected, especially negotiating the 2 different user IDs assigned to Facebook and Messenger user accounts.

Accomplishments that we're proud of

Getting a mostly functional machine learning model that can be interacted with live via a Facebook Messenger Chat Bot.

What we learned

Communication between many different components of the app; specifically the machine learning server, data parsing script, web server, and Facebook app.

What's next for Be Right Back

We would like to fully realize the goals of this project by training the model on a bigger data set and allowing more customization to specific users.

Built With

Share this project:

Updates