Over the past weekend, Semantics3 had a strong showing at the annual hackathon conducted by Sequoia Capital at Bengaluru. With four of our members invited to participate, we decided to enter as a single team.
The first (and possibly the most critical) issue was coming up with an idea for the hack. After considering the multiple themes for the hackathon and brainstorming over various ideas, the team decided to participate in the “Deep Learning” track.
Taking inspiration from recent research at the Facebook AI team, we decided to roll our own take on their bAbI datasets. The plan was to use this system to target the pain of support duty that most technical teams dread.
While we might get in to the detail of our hack in a future blogpost, I will try to summarize our submission with a few lines here.
We wanted to build a question answering system, which given a comprehensive dataset would be able to simplify the entire tech-support process.
Our development stack looked like most deep learning systems these days, making use of TensorFlow, via the useful Keras abstraction (we ended up implementing an end-to-end memory neural network model). Together with some quick front-end work using React, we were able to roll a considerable demo within the duration of the hackathon.
After many hours of sustained effort (filled with its own share of snafus), we finally managed to put together a working system — which was trained to answer support questions about the hackathon itself.
Without further adieu, here is the final pitch deck that we came up.
for those interested in the specifics, we’ll try to go into the details of our submission in a later post