Everyone has written a bad email before - one that doesn't make your intentions clear or conveys the wrong tone. This isn't just an embarrassing moment - it's a problem that costs billions per year in lost productivity. What if you could use AI to give you meaningful feedback before you hit send?
Communication extends beyond email. Conference calls are another common tool, it can be difficult to interpret critical nonverbal cues in a digital meeting. Collabra has developed a way to gauge emotions in a video call so that the host can easily see how engaged the participants are and adjust accordingly.
What it does
Collabra Email uses AI to put meaning back into email. It automatically parses out action items so that every recipient knows exactly what you expect from them. It suggests improvements and gives warnings about pitfalls in your email before you hit send, and integrates seamlessly with your existing Microsoft accounts.
Collabra Emotion displays an overlay over a video feed, identifying emotions expressed by each face detected in the frame.
How I built it
Collabra runs on a Python/Flask stack.
Sentiment analysis is performed using a combination of Microsoft Cognitive Services API and VADER Sentiment Analysis (via NLTK). Identifying action items is performed using a Naive Bayes classifier trained on a cleaned and hand-labelled subset of the Enron email dataset.
Video emotion detection is achieved using the Microsoft Cognitive Services API and OpenCV.
Challenges I ran into
Feature selection for action item identification, wrangling OAuth, dealing with Microsoft's strange Python documentation, asynchronous Python programming
Accomplishments that I'm proud of
Learning to clean datasets for machine learning to eliminate bias, not giving up
What's next for Collabra
Collabra Email is currently only a tool for reviewing emails before they're sent. The backend can easily be adapted to serve email recipients to automatically add upcoming events to their Outlook Calendar, evaluate urgency of to-do items and add them to Planner, and so on.
Collabra Emotion currently has a very low frame rate and can be optimized in various ways to improve that.