Meet Cindy, a machine learning engineer at Sentropy
Meet Cindy Wang, a machine learning engineer at Sentropy. Cindy talks about her passion for using machine learning to detect hate speech, her conviction that this technology will help solve the problem of abuse online, and the process of building it from scratch.
Disclaimer: An image in this post includes obscene language pertaining to hate speech.
What did you do before Sentropy?
I did my undergraduate and my master's at Stanford. Both were in computer science. So I doubled up and spent a lot of time in Palo Alto! During my undergraduate degree, I focused on theory, and during my master's I focused on artificial intelligence. I did my thesis project on hate speech detection using machine learning.
What do you do at Sentropy?
I’m a machine learning engineer at Sentropy. As part of the ML team, we work on everything that comes into and goes out of our models. Not only are we responsible for building the core model architecture but we also touch all the pieces that interact with that. That includes things like data ingestion, data pipelines, data preprocessing, building the models themselves, and serving those models in production. We work really closely with the data team and the services team to make sure that our models are high quality and can be served efficiently.
What problem is Sentropy solving?
In a nutshell, our founders saw this problem of abuse online and really wanted to make it better. So they gathered the best people they could find in order to create a solution.
Before Sentropy, I had worked on trust and safety problems both in industry and academia, and I realized that the solutions a lot of online platforms had for tackling online abuse were very elementary. Many times it was simply a list of bad words.
But the expertise to build something better from a technological standpoint exists. When I came across Sentropy I quickly realized that this team was pulling together experts across T&S and machine learning to build technology that would solve this problem. So it’s a matter of matching the solution with the people who have the capability to solve it. That’s why we’re here. We have a group of people who are all very motivated and invested in solving this problem. And we believe that we can build a better solution than the ones that already exist.
Could you describe an aspect of Sentropy’s culture that you’ve experienced during your time at the company?
We definitely have a culture that expects and encourages everyone to work cross-functionally with teams and projects across the company. Regardless of your role or title, your opinion is valued. Because of this, every person has insight into every area of the business and has their fingers in a lot of different pies, so to speak. You get to see a lot of the landscape — the whole tech stack and business side. For example, I feel comfortable making a suggestion on the product side or proposing a lead to the sales team. As a result we are a very closely-knit team.
On the machine learning team, we have a lot of channels for our coworkers from different teams to give us feedback on our models. We have an internal API for anyone at Sentropy to test and dogfood our product. We even recently built an internal tool that provides visual explanations of model predictions on a specific piece of content. It highlights how strongly different parts of an input influence our models’ predictions.
This lets people outside the machine learning team understand what’s going on under the hood with our models. It’s been very helpful for our data, product, and communications teams. They better understand our models and then come back to us with questions that help us improve the technology. How can we better build a model for this definition? What types of data are we missing?
How does diversity, equity, and inclusion play a role at Sentropy?
It goes without saying that this is one of our core values. It’s inherent to our mission of protecting people across the internet. It’s important to have the types of people that we are building for represented on our team.
In terms of my personal experience, I feel lucky to have had representation from day one. I did internships at lots of different companies before Sentropy and I had never had a woman as my manager. At Sentropy, I met Michele, who’s an expert in NLP (natural language processing) and someone who inspired me. She’s really thoughtful about this problem space and is deeply qualified to build a solution. She was also the first woman CTO I’d ever had the chance to potentially work with. This was really important to me and a major reason that I joined. At that point, the entire machine learning team was women, which isn’t something I’d ever expected or could even hope for in the past. Before this, I’d often had the experience of being the “token woman” in the room.
Treating diversity and inclusion as a first-order goal helps us build a better company and a better product. For the type of company that we’re building and the type of product that we want to put out into the world — it’s absolutely critical.
Why did you join Sentropy?
The reason that I decided to join was first and foremost because our mission is something that I care a lot about. Both personally and from an academic and technical perspective. On the technical side, it’s something that I had already spent a lot of time thinking about, had done academic research in, and had worked on at other companies.
But I wouldn’t have joined just any company that was working on abuse detection online. The team that we’ve put together is not only extremely strong in terms of the skills and capacity to solve this problem, it’s also really passionate and thoughtful about building useful tools and being mindful about how our product impacts real people. That means everyone from the users of the platforms we serve, to content moderators, to those annotating the data that goes into our models. I have confidence that our team and our approach can make a real difference in the way online abuse is handled.
How does Sentropy’s mission play a role in your day-to-day work?
Since we’re trying to build something in a space that doesn’t have an established playbook, we’re often brainstorming and designing things from scratch. Throughout this process, it’s been clear that we have a very thoughtful and skilled group of people here, and that our goal is first and foremost to build the right thing the right way. With ML products it’s easy to get mired in optimizing for, say, top-level metrics on a toy dataset. Instead, we ask ourselves how much abuse are we going to be able to protect people from online? How can we improve the health of a platform most effectively? How much can we improve someone’s experience online? That’s how we determine success. Whether it’s building a super accurate model or whether it’s building a, a platform that allows moderators to be more efficient. Our goal at the end of the day is to make the internet a safer place. As much of a catchphrase or motto as that sounds — it’s true.
We’re sprinting towards that goal because we believe that there is a technological solution to this problem of hate and abuse online. There’s a way to detect and mitigate it. And we believe that it’s in the solution that we’ve built. We’ve brought together a group of people with very diverse and deep experience and expertise. Whether it’s iterating on a new architecture with the ML team, finding new data sources with the data team, or improving our services latency there’s this undercurrent of a higher level goal that runs true for everyone in every one of these discussions.
Learn more about Cindy’s work. Read:
- Her blog series about addressing bias in machine learning classification tasks.
- And her post about building content moderation at scale with human-centered machine learning.
