"A lot of people in academia are not very good at software engineering," says Kenny Daniel, co-founder and chief technology officer of cloud computing startup Algorithmia. "I always had more of the software engineering bent."
That, in a nutshell, is some of what makes six-year-old, Seattle-based Algorithmia uniquely focused in a world over-run with machine learning offerings.
Amazon, Microsoft, Google, IBM, Salesforce, and other large companies have for some time been offering cut-and-paste machine learning in their cloud services. Why would you want to stray to a small, young company?
No reason, unless that startup had a particular knack for hands-on support of machine learning.
That's the premise of Daniel's firm, founded with Diego Oppenheimer, a graduate of Carnegie Mellon and a veteran of Microsoft. The two became best friends in undergrad at CMU, and when Oppenheimer went to industry, Daniel went to pursue a PhD in machine learning at USC. While researching ML, Daniel realized he wanted to build things more than he wanted to just theorize.
"I had the idea for Algorithmia in grad school," Daniel recalled in an interview with ZDNet. "I saw the struggle of getting the work out into the real world; my colleagues and I were developing state-of-the-art [machine learning] models, but not really getting them adopted in the real world the way we wanted."
He dropped out of USC and hooked up with Oppenheimer to found the company. Oppenheimer had seen from the industry side that even for large companies such as Microsoft, there was a struggle to get enough talent to get things deployed and in production.
The duo initially set out to create an App Store for machine learning, a marketplace in which people could buy and sell ML models, or programs. They got seed funding from venture firm Madrona Ventures, and took up residence in Seattle's Pike Place. "There's a tremendous amount of ML talent out here, and the rents are not as crazy" as Silicon Valley, he explained.
Their intent was to match up consumers of machine learning, companies that wanted the models, with developers. But Daniel noticed something was breaking down. The majority of customers using the service were consuming machine learning from their own teams. There was little transaction volume because companies were just trying to get stuff to work.
"We said, okay, there's something else going on here: people don't have a great way of turning their models into scalable, production-ready APIs that are highly available and resilient," he recalled having realized.
"A lot of these companies would have data scientists building models in Jupyter on their laptop, and not really having a good way to hook them up to a million iOS apps that are trying to recognize images, or a back-end data pipeline that's trying to process terabytes of data a day."
There was, in other words, "a gap there in software engineering." And so the business shifted from a focus on a marketplace to a focus on providing the infrastructure to make customers' machine learning models scale up.
The company had to solve a lot of the multi-tenant challenges that were fundamental limitations, long before those techniques became mainstream with the big cloud platforms.
"We were running functions before AWS Lambda," says Daniel, referring to Amazon's server-less offering.
Problems such as, "How do you manage GPUs, because GPUs were not built for this kind of thing, they were built to make games run fast, not for multi-tenant users to run jobs on them."
Daniel and Oppenheimer started meeting with big financial and insurance firms, to discuss solving their deployment problems. Training a machine learning model might be fine on AWS. But when it came time to make predictions with the trained model, to put it into production for a high volume of requests, companies were running into issues.
The companies wanted their own instances of their machine learning models in virtual private clouds, on AWS or Azure, with the ability to have dedicated customer support, metrics, management and monitoring.
That lead to the creation of an Algorithmia Enterprise service in 2016. That was made possible by fresh capital, an infusion of $10.5 million from Gradient Ventures, Google's AI investment operation, followed by a $25 million round last summer. In total. Algorithmia has received $37.9 million in funding.
Today, the company has seven-figure deals with large institutions, most of it for running private deployments. You could get something like what Algorithmia offers by using Amazon's SageMaker, for example. But SageMaker is all about using only Amazon's resources. The appeal with Algorithmia is that the deployments will run in multiple cloud facilities, wherever a customer needs machine learning to live.
"A number of these institutions need to have parity across wherever their data is," said Daniel. "You may have data on premise, or maybe you did acquisitions, and things are across multiple clouds; being able to have parity across those is one of the reasons people choose Algorithmia."
Amazon and other cloud giants each tout their offerings as end-to-end services, said Daniel. But that runs counter to reality, which is that there is a soup composed of many technologies that need to be brought together to make ML work.
"In the history of software, there hasn't been a clear end-to-end, be-all winner," Daniel observed. "That's why GitHub, and GitLab, and Bitbucket and all these continue to exist, and there are different CI [continuous integration] systems, and Jenkins, and different deployment systems and different container systems."
"It takes a fair amount of expertise to wire all these things together."
There is some independent support for what Daniel claims. Gartner analyst Arun Chandrasekaran puts Algorithmia in a basket that he calls "ModelOps." The application "life cycle" of artificial intelligence programs,
Chandrasekaran told ZDNet, is different from that of traditional applications, "due to the sheer complexity and dynamism of the environment."
"Most organizations underestimate how long it will take to move AI and ML projects into production."
Chandrasekaran predicts the market for ModelOps will expand as more and more companies try to deploy AI and run up against the practical hurdles.
While there is the risk that cloud operators will subsume some of what Algorithmia offers, said Chandrasekaran, the need to deploy outside a single cloud supports the role of independent ModelOps vendors such as Algorithmia.
"AI deployments tend to be hybrid, both from the perspective of spanning multiple environments (on-premises, cloud) as well as the different AI techniques that customers may use," he told ZDNet.
Aside from cloud vendors, Algorithmia competitors include Datarobot, H20.ai, RapidMiner, Hydrosphere, Modelop and Seldon.
Some companies may go 100% AWS, conceded Daniel. And some customers may be fine with generic abilities of cloud vendors. For example, Amazon has made a lot of progress with text translation technology as a service, he noted.
But industry-specific, or vertical market machine learning, is something of a different story. One customer of Algorithmia, a large financial firm, needed to deploy an application for fraud detection. "It sounds crazy, but we had to figure out all this stuff of, how do we know this data over here is used to train this model? It's important because its an issue of their [the client's] liability."
The immediate priority for Algorithmia is a new product version called Teams that lets companies organize an invite-only, hosted gathering of those working on a particular model. It can stretch across multiple "federated" instances of a model, said Daniel. The pricing is by compute usage, so it's a pay-as-you-go option, versus the annual billing of the Enterprise version.
To Daniel, the gulf that he observed in academia between pure research and software engineering is the thing that has always shot down AI in past. The so-called "AI winter" periods over the decades were in large part a result of the practical obstacles, he believes.
"Those were periods when there was hype for AI and ML, and companies invested a lot of money," he said. "If companies are not getting the pay-off, if there's a lack of progress, we could be looking at another hype cycle."
By contrast, if more companies can be successful in deployment, it may lead to a flourishing of the kind of marketplace that he and Oppenheimer originally envisioned.
"It's like the Unix philosophy, these small things combining, that's the way that I see it," he said. "Ultimately, this will just enable all sorts of things, completely new scenarios, and that's incredibly valuable, things that we can make available in a free market of machine learning."