Miso

Why are there so few women in STEM? How can we promote female empowerment within STEM? As two of us are women, we have personally faced misogyny within the tech community—from being told women aren't fit for certain responsibilities, to our accomplishments being undermined, we wanted to diminish gender inequality in STEM.

We chose to focus on one of the roots of this problem: misogynistic or sexist comments in both professional and social settings.

Thus, we built Miso, an AI that detects sexism and toxicity in text messages and warns the sender of their behaviour. It is then up to them whether they still want to send it or not, but mind that messages which are filtered as at least 60% sexist or 80% toxic will be sent up to admin within the communication channel.

It can be integrated into Discord as a bot or used directly on our website. In a professional workspace, which Discord is increasingly being used as—especially for new startups—the consequences of toxic behaviour are even more serious.

Every employee in the company will have a profile which tracks every time a message is flagged as sexist or toxic. This tool helps HR determine inappropriate behaviour, measure the severity of it, and keep records to help them mediate workplace conflicts.

We hope to reduce misogyny within tech communities so that women feel more empowered and encouraged to join STEM!

We made our own custom machine learning model on Cohere. We made a classify model which categorises text inputs into the labels True (misogynistic) and False (non-misogynistic). To train it, we combined various databases which took comments from social media sites including Reddit and Twitter, and organised them into 2 columns: message text and their associated label. In addition to our custom model, we also implemented the Cohere Toxicity Detection API into our program and checked messages against both models.

Next, we developed a Discord Bot so that users can integrate this AI into their Discord Servers. We built it using Python, Discord API, and JSON. If someone sends a message which is determined to be 60% likely to be sexist or 80% likely to be toxic, then the bot would delete the message and send a warning text into the channel.

To log the message history of the server, we used Estuary. Whenever a message would be sent in the discord, a new text file would be created and uploaded to Estuary to be backed up. This text file contains the message id, content and author. The file is uploaded by calling the Estuary API and making a request to upload. Once the file uploads, the content ID is saved and the file on our end is deleted. Estuary allows us to save message logs that are problematic without the burden of storage on our end.

Our next problem was connecting our machine learning models to our front end website & Discord. The Discord was easy enough as both the machine learning model and the discord bot was coded in Python. We imported the model into our Discord bot and ran every message into both machine learning models to determine whether they were problematic or not.

The website was more difficult because it was made in React.JS so we needed an external app to connect the front end display and back end models. We created an API endpoint in our model page using FastAPI and created a server using Uvicorn. This allowed us to use the Fetch API to fetch the model's prediction of the inputted sample message.

As for our frontend, we developed a website to display Miso using React.JS, Javascript, HTML/CSS, and Figma.

Some challenges we ran into were figuring out how to connect our frontend to our backend, incorporate Estuary, train a custom machine learning model in Cohere, and handle certain libraries, such as JSON, for the first time.