While our team might have come from different corners of the country, with various experience in industry, and a fiery desire to debate whether tabs or spaces are superior, we all faced similar discomforts in our jobs: insensitivity.

Our time in college has shown us that despite the fact people's diverse backgrounds, everyone can achieve greatness. Nevertheless, workplace calls and water-cooler conversations are plagued with "microaggressions." A microaggression is a subtle indignity or offensive comment that a person communicates to a group. These subtle, yet hurtful comments lead to marginalization in the workplace, which, as studies have shown, can lead to anxiety and depression. Our team's mission was to tackle the unspoken fight on diversity and inclusion in the workplace.

Our inspiration came from this idea of impartial moderation: why is the marginalized employee's responsibility to take the burden of calling someone out? Pointing out these microaggressions can lead to the reinforcement of stereotypes, and thus, create lose-lose situations. We believe that if we can shift the responsibility, we can help create a more inclusive work environment, give equal footing for interviewees, and tackle marginalization in the workplace from the water-cooler up.

What it does


EquiBox is an IoT conference room companion, a speaker, and microphone that comes alive when meetings take place. It monitors different meeting members' sentiment levels by transcribing sound and running AI to detect for insults or non-inclusive behavior. If an insult is detected, EquiBox comes alive with a beep and a warning about micro-aggressions to impartially moderate an inclusive meeting environment. EquiBox sends live data to EquiTrack for further analysis.


EquiTalk is our custom integration with Twilio (a voice platform used for conference calls) to listen to multi-person phone calls to monitor language, transcribe the live conversation, and flag certain phrases that might be insulting. EquiTalk sends live data to EquiTrack for analysis.


EquiTrack is an enterprise analytics platform designed to allow HR departments to leverage the data created by EquiTalk and EquiBox to improve the overall work culture. EquiTrack provides real-time analysis of ongoing conference calls. The administrator can see not only the amount of micro-aggression that occur throughout the meeting but also the direct sentence that triggered the alert. The audio recordings of the conference calls are recorded as well, so administrators can playback the call to resolve discrepancies.

How we built it

The LevelSet backend consisted of several independent services. EquiTalk uses a Twilio integration to send call data and metadata to our audio server. Similarly, EquiBox uses Google's VoiceKit, along with Houndify's Speech to Text API, to parse the raw audio format. From there, the transcription of the meeting goes to our micro-aggression classifier (hosted on Google Cloud), which combines a BERT Transformer with an SVC to achieve 90% accuracy on our micro-aggression test set. The classified data then travels to the EquiTalk backend (hosted on Microsoft Azure), which stores the conversation and classification data to populate the dashboard.

Challenges we ran into

One of the biggest challenges that we ran into was creating the training set for the micro classifier. While there were plenty of data sets that including aggressive behavior in general, their examples lacked the subtlety that our model needed to learn. Our solution to this was to crowdsource and augment the set of the microaggressions. We sent a survey out to Stanford students on campus and compiled an extensive list of microaggressions, which allowed our classifier to achieve the accuracy that it did.

Accomplishments that we're proud of

We're very proud of the accuracy we were able to achieve with our classifier. By using the BERT transformer, our model was able to classify micro-aggressions using only the handful of examples that we collected. While most DNN models required thousands of samples to achieve high accuracy, our micro-aggression dataset consisted of less than 100 possible micro-aggressions.

Additionally, we're proud of our ability to integrate all of the platforms and systems that were required to support the LevelSet suite. Coordinating multiple deployments and connecting several different APIs was definitely a challenge, and we're proud of the outcome.

What we learned

  • By definition, micro-aggressions are almost intangible social nuances picked up by humans. With minimal training data, it is tough to refine our model for classifying these micro-aggressions.

  • Audio processing at scale can lead to several complications. Each of the services that use audio had different format specifications, and due to the decentralized nature of our backend infrastructure, merely sending the data over from service to service required additional effort as well. Ultimately, we settled on trying to handle the audio as upstream as we possibly could, thus eliminating the complication from the rest of the pipeline.

  • The integration of several independent systems can lead to unexpected bugs. Because of the dependencies, it was hard to unit test the services ahead of time. Since the only way to make sure that everything was working was with an end-to-end test, a lot of bugs didn't arise until the very end of the hackathon.

What's next for LevelSuite

We will continue to refine our micro-classifier to use tone classification as an input. Additionally, we will integrate the EquiTalk platform into more offline channels like Slack and email. With a longer horizon, we aim to improve equality in the workplace in all stages of employment, from the interview to the exit interview. We want to expand from conference calls to all workplace communication, and we want to create new strategies to inform and disincentivize exclusive behavior. We wish to LevelSet to level the playing the field in the workplace, and we believe that these next steps will help us achieve that.

Share this project: