Tecton, the machine learning (ML) “feature platform” company founded by the creators of Uber’s Michelangelo ML platform, today announced version 0.6 of its product. The update allows users to “build production-ready features right into their notebooks and get them up and running in minutes,” said Mike Del Balso, Tecton co-founder and CEO.
I spoke with Del Balso on a Zoom call to find out exactly what an ML service platform is—and what it’s typically used for in enterprises. Also on the call was Gaetan Castelein, head of marketing at Tecton.
Table of Contents
What is the function and what does it do?
“When you think about a machine learning application, there are two parts to it,” Del Balso said. “There is a model that will eventually predict it. But then that model […] you need to accept some data inputs – these data inputs are the characteristics. And these features contain all the relevant information about the world that you need to know right now to make a good prediction.”
An example of this feature is data about how busy roads are for an Uber ride. Or is there rush hour traffic? Both datasets would be “services” for an ML application.
Graphic via Tecton
In fact, Del Balso and Tecton co-founder Kevin Stumpf (CTO) worked at Uber with the idea of a “feature platform”. According to Tecton’s About page, the pair built the Michelangelo ML platform at Uber, which was “instrumental in helping Uber scale to 1,000. [ML] supporting a wide range of models coming into production within a few years, from real-time pricing to fraud detection and ETA prediction.”
They quickly realized that a service platform could be used in any ML workload that included what Del Balso called “real-time production machine learning.” Before Uber, Del Balso worked at Google “on the machine learning that powers Google’s advertising systems.” Additional use cases for Tecton’s technology include recommendation systems, real-time dynamic pricing, and payment system fraud detection.
Define and run a function
Tecton’s primary users are data scientists or engineers, and this requires code to define a feature. According to the documentation, Tecton’s services are “defined as views relative to a Python, SQL, Snowpark, or PySpark data source.”
“It’s not a no-code platform or anything like that,” Del Balso confirmed. “When you use a service platform, you define the code, you define the transformations that take the raw data from your business and turn it into the data — the features — that the model uses to make predictions.”
Tecton concept diagram (via Tecton); click to see it in full
After defining functions through code, the service platform “handles every aspect of these data streams at every stage of the machine learning lifecycle,” he said.
This includes calculating and updating the data itself, all during the process.
The service platform “constantly calculates the latest values of these signals so that the model always has the most relevant information. [in order] to make the most accurate prediction,” he explained.
Bridging the developer and production
Because machine learning in applications is still relatively new at the company, Tecton users often have a mix of skills.
“We’re in this interesting area of the industry where […] machine learning teams look very different between companies,” said Del Balso. “So our target is the people who build your machine learning application. It could be a data scientist who doesn’t have a manufacturing engineering background, but very often a company has an engineer who has a manufacturing engineering background but may not really be an expert in data science.”
Where there have been problems in the past is the “wall” between the development environment and the production environment. Data scientists, in particular, usually do not have experience in deploying an application to production. Tecton aims to solve this, said Del Balso.
“It’s these two different worlds, data scientists and engineers didn’t know how to work with each other – during development, let alone ongoing operations. And the value of the feature platform is to break down that wall, making it easier. It provides a centralized, single way for data scientists to define these features in their development workflows, with essentially no additional tasks for production.”
With version 0.6 of its platform, Tecton claims to have integrated the functionality workflow into a data scientist’s existing notebook tools. According to Del Balso, this removes the barriers that prevent data scientists from easily going into production.
“Now you don’t even have to leave your data science tools,” he said. “You can use the same Jupyter Notebook. You can use the same data science environment you built or are used to. So the experience is much closer to what they are [data scientists] they love and feel good. And it allows us to bring development and production environments and experiences closer than ever before.”
AI in the Enterprise
While generative artificial intelligence continues to grab all the headlines (OpenAI released GPT-4 this week), it’s equally interesting to track how AI and machine learning enter the world of enterprise IT. Just as we saw the DevOps revolution in the late 2000s and 2010s after the advent of cloud computing, we are now seeing an “MLOps” (for lack of a better term) revolution in the early 2020s as AI takes hold.
Overall, Tecton is another example of the growing range of AI tools that are becoming increasingly important in the business environment.
Richard MacManus is the senior editor of The New Stack and writes about web and app development trends. He previously founded ReadWriteWeb in 2003 and built it into one of the most influential technology news sites in the world. From the beginning…