Sentiment Seer

Inspiration

We think Jira is a super useful tool, but it can be overwhelming, too. We are trying to make working with Jira even more managable and enjoyable.

What it does

Extracts sentiments from comments and uses closed issues to train our ML model that generates predictions for ticket outcomes.

How we built it

It is partly trigger-based: We utilize the Jira comment event to extract moods the and issue updated event to trigger training our model on newly closed issued.
We use the Storage API to store the model weights and the Property API to store extracted comment moods.
We use a project settings page as GUI for the user to view predictions for the issues in their project.

Challenges we ran into

The 25s limit for normal function calls was challenging when training and predicting our model. Collecting even a couple issues and corresponding attributes and train them can take that long. We solved it by only training small batches of issues and using JQL paging when predicting issues. There's room for improvement, though.
For the ML part, we struggled with training a model of our own. Classic training on huge data sets proved unpractical so we were happy to find the online learning approach.

Accomplishments that we're proud of

Implementing the passive/aggressive algorithm from scratch and training/predicting the whole model on Forge architecture.

What we learned

This was the first time we touched ML, so we had to learn a lot about it.

What's next for Sentiment Seer

We wanted to build something useful and we are quite happy with it. We are thinking about expanding the use case and might even put it in the marketplace.