Back in 2019, Kannan Kothandaraman helped to start Selector AI after spending years in the networking trenches at big vendors including Juniper and Cisco.
Since the beginning, the vision for Selector AI is to build a technology that helps networking professionals better operate and optimize network infrastructure and services. The company’s founding team was inspired by what it learned while working with hyperscalers. They noticed how these large companies are able to use data from multiple domains to make data-driven decisions for their network operations, rather than letting the network dictate how the business should be run. The Selector AI founders saw an opportunity to bring this data-driven approach to non-hyperscaler companies and started building a solution alongside early design partners. It’s an approach that has continued to grow and scale over the years.
In September, the company debuted its network language model for netops, bringing the power of generative AI to the platform. This week the company announced the latest milestone in its evolution, raising a $33 million Series B round of funding. Total funding to date for Selector AI now tops $66 million. The new funding will be used to help the company grow sales and marketing as well technical innovation.
“One of our key differentiators is that any type of data that an operations team is using for a network or an infrastructure, we are able to consume that,” Kothandaraman told Network World. “There’s a lot of proprietary data that every team uses.”
Overcoming challenges: From dirty data to deterministic outputs
The journey to Selector AI’s current success was not without its challenges. In the early days, the team faced difficulties in raising their first round of funding, as the COVID-19 pandemic created a challenging fundraising environment.
Additionally, the team quickly learned that the reality of working with customer data was far from what they had initially envisioned.
“Data is very dirty, right? You know, you have to make do with whatever is available in a lot of cases,” Kothandaraman said. “You’ll have to augment a lot of it with your own data, and nobody is willing to share data with each other.”
Another challenge that emerged was dealing with the limitations and potential of different forms of AI. “AI and ML are inherently probabilistic, whereas operational teams do not want any probabilistic answers. It has to be very deterministic,” he said.
A deterministic response is one that is accurate and consistent time after time, and when dealing with operations, it’s absolutely essential to success. Kothandaraman noted that for network operations teams, when they get trouble tickets and alerts, the information has to be actionable. “We can’t send them on a probabilistic wild goose chase,” he said.
Solving the network noise problem
A key benefit of the Selector AI platform is its ability to reduce the overwhelming amount of alerts that plague network operations teams.
“Some customers have like 30-40 different tools that they are using, each of them starts to complain whenever there is an issue,” Kothandaraman noted. “The number one criteria that we are judged on is noise reduction.”
The platform achieves this through sophisticated data integration and relationship mapping. Kothandaraman explained that Selector AI uses both operational and relationship data. That data is what is fed into the machine learning engine to help identity issues.
“If I have a problem, all these different tools are complaining, but where is the root of that problem? That’s what we focus on,” he said.
Looking ahead: Digital twins and expanded language models
As Selector AI sets its sights on 2025 and beyond, Kothandaraman outlined two key areas of focus for the company’s technology development.
The first area is data-driven digital twins. The concept of a digital twin is some form of replica of an existing network that can be used in a virtual way to identify issues in a real network without impacting the production network. Kothandaraman said that Selector AI will be focused on a data-driven, rather than a configuration-driven approach to digital twins.
The second area of investment will be in expanded language models for netops. “Network and infrastructure is what we are focused on,” Kothandaraman said. “How do we combine real time data with language models? That is not a solved problem at this point, so we are looking to solve that for network and infrastructure.”
Source:: Network World