Angular 2: a nightmare as always
Machine learning: broken
Sometimes, you're busy. You don't have time to respond to messages. Maybe your grandma has sent you a link to a cake recipe she found. It would be rude to ignore her, but you're very busy indeed!
What it does
Hullo taps into your conversation data from facebook messenger, google hangouts and the like to train a computer to talk like you. When you're not available, Hullo can step in and respond to simple queries.
How we built it
We downloaded corpuses of message data from various services, cleaned and regularised them, fed them through some machine learning, built an actor out of the learned model and attached it to a lightweight webserver. Then we built a web app to chat with it. And plugged it into Alexa.
Challenges we ran into
Machine learning, it turns out, is quite hard. Our original plan to use a differentiable neural computer fell through after many hours of experimentation. We're currently implementing an alternative model using stacked LSTMs.
Accomplishments that we're proud of
Finally getting Alexa to work. Ish.
What we learned
Before you go to a hackathon, become familiar with your tools.
What's next for hullo
- Mimic your voice?
- Plug into more services
- More and better ML
- More convenient import of chat data