A Look Inside FeatureByte’s Founding Story and Culture
Every startup has a story. FeatureByte’s was born out of a pain point our two founders, Razi and Xavier, experienced first hand. Most companies building AI tools and tech are focused on models. But, there’s not enough emphasis on the data that goes into the models to do predictions and train them. That’s where the value comes in the form of major performance improvements.
I recently sat down with our two founders to interview them about the motivations behind starting FeatureByte, and the culture that’s driving us toward the important goal of simplifying and industrializing feature creation, deployment and management. Here’s what they had to say.
How did you two meet and decide to start FeatureByte?
Razi: Both of us were early employees at DataRobot. Even though Xavier was in Singapore and I was in Boston, we got to know each other and became friends. The overall idea for FeatureByte has been in the back of our heads for some time.
AI tools focus too heavily on achieving incremental performance improvements in models by tuning hyperparameters, trying out different algorithms, and other methods. Feature engineering and management (FEM) is one of the last remaining hurdles for many enterprises to be able to scale their AI efforts.
As a result, the potential for AI and ML in enterprises is being capped. There’s a huge gap between raw data and running accurate machine learning models in production.We want to help enterprises close that gap and deploy AI at scale, in more areas of the enterprise. By simplifying and industrializing AI data management, we can do that.
How do existing tools fall short when it comes to feature engineering and management?
Xavier: The modern data stack for BI doesn’t yet exist for AI. Today, some people think BI workflows can be used for AI because the tools exist and are already mature. But they’re designed for BI, which is often focused on answering very specific questions instead of experimentation.
Razi: We’re making AI more data centric instead of model-centric. We’ve seen different companies and even some open source projects focused on building feature stores. Even though the tech around feature stores is necessary, it’s not sufficient enough to allow for the scalable and extensive AI data management needed by data scientists.
Xavier: We could see that some elements were missing from feature stores. One example is feature transforms. The modern data stack for BI has something equivalent, however, a feature transform is complex. There’s no good solution for data scientists for data transformation.
Code is complex due to time-awareness and other requirements. Data scientists write features with certain tools, but what they write cannot be used in production. Data engineers use a different solution to transform code at the end, resulting in inconsistencies. We wanted to give data scientists full control over the feature transform process.
Another example is the feature catalog. While a data catalog for BI exists, we saw an opportunity to do things much better for AI. Finally, feature stores don’t address feature development. We wanted to give data scientists the ability to write production-ready features, and codify or declare features using tools that they’re familiar with.
Was the idea for FeatureByte born out of a specific experience you had?
Xavier: I love the creative aspect of building features. I like to understand the business and build better features based on what I know about their needs. I found that when I was building features, it was complex to get them into production.
One way to address that was to develop strong skills in data engineering, but this would impact my ability to understand the business and develop powerful features. FeatureByte was created out of a need to unleash creativity of data scientists. They need to invest their time in understanding the business, and formulating solutions to problems.
With FeatureByte, when they create a feature, they can deploy into production in no time. They don’t have to rely on different teams to make things happen.
Razi: In our time at DataRobot and after that, we’ve worked with hundreds of companies implementing AI for building and deploying models. FeatureByte is based on first hand experience with the AI data management challenges companies face.
How are you bringing the idea for FeatureByte to life as a product?
Razi: We are fortunate to build a very strong team. Our key differentiators are our team and market opportunity. Our team members get the opportunity to work with world-class data scientists. The leadership team has been on the business side of successful unicorns and large companies alike. Every single one of our team members has an opportunity to help the company grow, and see their contributions as a part of something much bigger.
What are some things that are important to you when it comes to building a company culture?
Xavier: At our core, we’re not afraid of challenges. We accept them well and see them as opportunities. Our team is motivated to be excited about the unknown.
Razi: We want to build an inclusive and friendly culture, which comes with a lot of transparency and active communication. Our team is distributed in the U.S. and Singapore.
Without establishing transparency and trust it would be hard to have a successful team culture. We’re constantly aligning teams together so we are rowing in the same direction.
We’re also open about our roadmap and explaining how we make choices within the organization. That comes down to both engineering decisions, and higher-level details about the company, product, and market.
We believe team members should be able to feel comfortable expressing individual challenges, whether it’s personal or professional. Regardless of their level of expertise, a team member can feel comfortable expressing what they don’t know and flagging where they need additional support. This type of dialogue is important for personal growth. We’re creating a safe environment where people can grow, contribute, and learn from each other. Junior employees love the ability to stretch and try new things, with a support network in place.
Why did you decide to release the core as Open Source as part of launching FeatureByte?
Xavier: We wanted to give the power of FeatureByte to every data scientist and engineer out there, so they can focus on creativity, experimentation and problem solving, rather than data pipelines and plumbing. This is also a way to start codifying feature engineering best practices, so we can innovate much faster as an industry. Just imagine the impact when everyone can create state-of-the-art features and deploy pipelines in minutes, with just a few lines of code!
This is just the beginning. Our goal is to truly democratize and industrialize feature creation and deployment. We’re creating a feature engineering community to contribute and extend the core, so that more industries and even smaller companies can benefit from AI.