Union.ai raises $10M to simplify AI and ML workflow orchestration

Union.ai, a startup emerging from stealth with a commercial version of the open source AI orchestration platform Flyte, today announced that it raised $10 million in a round contributed by NEA and “select” angel investors. CEO Ketan Umare says that the proceeds will be put toward supporting the Flyte community by “improving the accessibility, performance and reliability of Flyte” and broadening the array of systems that Flyte integrates with.

While companies find AI’s predictive power alluring, particularly on the data analytics side of the organization, achieving meaningful results with AI often proves to be a challenge. It’s true that AI can help to project revenue, for example, by identifying trends in buying and selling. But implementing and maintaining the data pipelines necessary to keep AI systems from drifting to inaccuracy can require substantial technical resources.

That’s where Flyte comes in — a platform for programming and processing concurrent AI and data analytics workflows. Union’s team, including Umare, helped to build Flyte while at Lyft, where it was used to help create a system to calculate the estimated time of arrival (ETA) for drivers to get from point A to point B.

“[Union’s] founders first met at Lyft, where we joined the team responsible for calculating the ETA for a Lyft driver to get from point A to point B,” Umare told TechCrunch via email. “Searching for the right solution led the team deep into machine learning techniques, which came with requirements to use large amounts of data and deliver robust models to production consistently … The techniques used were platformized, and the solution was used widely at Lyft.”

Lyft contributed Flyte to open source in 2020, granting the trademark to the Linux Foundation a year later. That’s when Union’s team saw an opportunity to layer paid services on top of the project in the cloud.

“A managed version of Flyte, called Union Cloud, will allow smaller teams and organizations to use the power of Flyte without the need to staff up on infrastructure teams,” Umare continued. “We [founded Union] because we believe that machine learning and data workflows are fundamentally different from software deployments. This is because software is more precise with a slower lifecycle while machine learning and data workflows start off being experimental and may need to be quickly productionized.”

Taking Flyte

Umare and Union’s other co-founders, Haytham Abuelfutuh and George Snelling, all have deep backgrounds in the tech industry. Prior to joining Lyft, Umare was a senior software engineer at Amazon and a principal engineer at Oracle, where he led development of a block storage product for an infrastructure-as-a-service and bare metal offering. Abuelfutuh spent seven years as an engineer at Microsoft and three as a developer at Google, where he helped to ship an internal software library for first-party apps including Google Photos. Snelling — also a Microsoft veteran — co-founded several startups (Westside, LabKey and Patchr) and spent time at Salesforce as a senior director of engineering.

With Union Cloud — the launch of which coincides with the release of Flyte version 1.0 — Umare says the goal is to reduce (and ideally eliminate) the unwieldy infrastructure that can crop up in data science projects and hamstring development. At their worst, messy abstractions can necessitate rebuilding infrastructure to deploy AI to production, Umare points out — negatively affecting the potential return on investment.

According to a 2021 Wakefield Research report, enterprise data engineers spend nearly half their time building and maintaining data pipelines. Sixty-nine percent of respondents to the survey — mainly data engineers — said that business outcomes would improve if their teams could contribute more to business decisions and spend less time on manual pipeline management.

“Production machine learning is still in its infancy at the moment, especially at companies outside big tech. Thus, most companies start off with DIY — that is our primary competition,” Umare said. “We took a radically different, first-principles approach to defining what a workflow means for machine learning and data scientists. We started with a goal to minimize human errors and try to help predict problems ahead of time [and worked] closely with extremely sophisticated and a diverse set of partners like Spotify, Gojek and Freenome [to help] refine the solution.”

Union Cloud inherits all of Flyte’s characteristics and capabilities, including connectors between computation back ends that record all changes to an AI pipeline. Union Cloud also stores a history of all a pipeline’s executions and provides a dashboard, command-line interface and API to interact with the computations.

Union Cloud — and Flyte — define workflows as multiple tasks. Workflows and tasks can be written in any programming language and stay on-premises, as does data moving through those components.

Cloud advantage

So what’s the value add with Union Cloud? Umare says that it adds “agility, reproducibility, and security” to Flyte by centralizing infrastructure management and maintaining “high” privacy and compliance standards. “Our products are built with zero-trust principles in mind and thus our users can use [it] to build a self-serve platform that still maintains high security standards,” he continued. “Data science is very academic, which directly affects machine learning. There is a lot of fantastic research and literature that is available in academia, which is hard to productionize. We need to bridge both these worlds in a structured and repeatable way.”

Umare also sees Union Cloud as a way to reduce the cost of developing new products and systems in a way that the open source Flyte project can’t accomplish. While he concedes that similar efforts from other vendors exist, like AWS Sagemaker, he believes that they fail to integrate well with the rest of the data science ecosystem.

“We have been at this problem for over five years, refining our solution and iterating based on real-world feedback and requirements,” Umare said. “The machine learning sector is already large and growing within traditional companies as well. We view growth potential to not be limited by the size of the current demand however, but rather by the experience we can deliver, which is why we’ve focused purely on customer success and open source adoption. This will lead to revenue growth in the near future.”

On the topic of growth, Union plans to double its 20-person headcount by the end of the year as it focuses on product buildout. Umare didn’t have statistics to share on Union Cloud interest or uptake, but reiterated that “thousands” of users across companies such as Lyft, Spotify, Toyota subsidiary Woven Planet, and biotech and finance brands have adopted Flyte.