Number station, a startup that uses large language models (LLMs) as the muse of its data analytics platform, today launches its first cloud-based product: the aptly named Numbers Station Cloud, now in early access. This service allows virtually any user in an organization to investigate their internal data through Numbers Station's chat interface.
Several similar tools give attention to translating natural language queries right into a database language similar to SQL. However, the Numbers Station team argues that this approach has limitations, partly since the generic LLM lacks an understanding of how a specific company operates, the way it structures its data, and the way people inside the company consult with specific data objects.
Photo credit: Number station
As co-founder and CEO of Numbers Station Chris Aberger He told me he's type of bored with talking about how the service allows users to “chat with their data” because there's a lot noise about it. “But at the upper level, executives and non-technical users have questions they wish to ask after which get answers through these classic structured data sources. That’s really where things have come to,” he told me. “There’s quite a lot of data modeling, data exploration around these basic models and huge language models to make them work.”
For Numbers Station, this meant investing quite a lot of technical resources into constructing what the corporate calls its semantic catalog. This catalog is basically an routinely curated source of an organization's metrics and definitions. This catalog is restricted to every company (and isn’t shared between them). Aberger described the catalog as “a disgusting thing” that ensures, for instance, that the model’s definition of “recurring revenue” matches the corporate’s use of that term.
While Numbers Station's platform is built on plenty of very specialized LLMs and machine learning models, it is that this catalog that ties all of it together. As co-founder and chief scientist of Numbers Station Ines Chami I used to be told that the team originally underestimated the challenge of constructing this a part of the platform.
“It goes back to classical (machine learning) and classical data engineering: How do I create a representation of data that the model can actually use to reply these questions,” she told me. “Because there’s no way a model goes to know all of those metrics and all the things that business users are asking about.” Because not even humans understand every query immediately, and the model has to show these vague questions into very specific questions. number stations' Research shows that its approach ends in significantly improved precision in comparison with more traditional text-to-SQL pipelines.
Photo credit: Number station
Although the corporate is launching this chat service today, the general vision is significantly larger.
“Essentially we’re constructing an AI platform for analytics,” Aberger said. “This is certainly one of the applications (…). There's a bigger, broader project that we're still working on as an organization that addresses plenty of different data issues that come up here. Examples include: How do I enrich my data with third-party data? Sources? How do I apply a few of these more classic algorithms like fuzzy matching etc.? There are almost infinite spokes from which to construct on this platform.”
The company has already secured several Fortune 500 clients, including global real estate services firm Jones Lang LaSalle. “Numbers Station is a frontrunner in enterprise AI for structured data,” said Sharad Rastogi, CEO of Work Dynamics Technology at Jones Lang LaSalle. “We are impressed with Numbers Station’s trustworthy and interesting platform. It continually learns as we use it, enabling our data teams to find and test hypotheses to drive impactful business outcomes.”
