A new startup called Arize AI is building what it calls a real-time analytics platform for “observability” in artificial intelligence and machine learning.
The company is led by CEO Jason Lopatecki, who has also served as chief strategy officer and chief innovation officer at TubeMogul, the video ad company acquired by Adobe. TubeMogul’s co-founder and former CEO Brett Wilson is an investor and board member.
And it has already made an acquisition: a Y Combinator -backed startup called Monitor ML. The entire Monitor ML team is joining Arize, and its CEO Aparna Dhinakaran (who previously built machine learning infrastructure at Uber) is becoming Arize’s co-founder and chief product officer.
Lopatecki and Dhinakaran said that even when they were leading two separate startups, they were trying to solve similar problems — problems that they both saw at big companies.
“Businesses are deploying these complex models that are hard to understand, they’re not easy to troubleshoot or debug,” Lopatecki said. So if an AI or ML model isn’t delivering the desired results, “The state of the art today is: You file a ticket, the data scientist comes back with a complicated answer, everyone’s scratching their head, everyone hopes the problem’s gone away. As you push more and more models into the organization, that’s just not good enough.”
Similarly Dhinakaran said that at Uber, she saw her team spend a lot of time “answering the question, ‘Hey, is the model performing well?’ And diving into that model performance was really a tough problem.”
To solve it, she said, “The first phase is: How can we make it easier to get these real-time analytics and insights about your model straight to the people who are monitoring it in production, the data scientist or the product manager or engineering team?”
Lopatecki added that Arize AI is providing more than just “a metric that says it’s good or bad,” but rather a wide range of information that can help teams see how a model is performing — and if there are issues, whether those issues are with the data or with the model itself.
Besides giving companies a better handle on how their AI and ML models are doing, Lopatecki said this will also allow customers to make better use of their data scientists: “[You don’t want] the smallest, most expensive team troubleshooting and trying to explain whether it was a correct prediction or not … You want insights surfaced up [to other teams], so your head researcher is doing research, not explaining that research to the rest of the team.”
He compared Arize AI’s tools to Google Analytics, but added, “I don’t want to say it’s an executive dashboard, that’s not the right positioning of the platform. It’s an engineering product, similar to Splunk — it’s really for engineers, not the execs.”
Lopatecki also acknowledged that it can be tough to make sense of the AI and ML landscape right now (“I’m technical, I did EECS at Berkeley, I understand ML extremely well, but even I can be confused by some of the companies in this space”). He argued that while most other companies are trying to tackle the entire AI pipeline, “We’re really focusing on production.”