Featureform Inc., the developer of a feature store for artificial intelligence development, today said it has raised a $5.5 million seed round that brings its total funding to $8.1 million.

The company is addressing what it says is a mainly bespoke process by which companies transform raw data into signals to feed into large language models or machine learning models as inputs. Features are the input variables that an AI model uses to make predictions or decisions. The quality and selection of features directly influence the quality of the model’s performance.

Featureform has developed a unified framework that’s available under an open-source license with commercial extensions that developers can use to define, manage and deploy features. The platform has built-in search, monitoring, orchestration and governance.

Founder and Chief Executive Simba Khadder likens it to Terraform, a widely used infrastructure-as-code platform that inspired Featureform’s name. Terraform enables developers and administrators to define and provision data center infrastructure using a declarative configuration language.

“I think of a feature as a definition,” he said. “We store the definition, run it for you on [data analytics platforms] Databricks or Snowflake, define your pipeline, define the metadata and versioning, make it shareable, monitor for model drift and make sure governance is properly used. We’re a full platform to define and manage features, not just a place to store them.”

Features are stored in a vector database, which is designed specifically for handling the sequence of numbers or coordinates that represent data in the multidimensional form used in machine learning. Data scientists can use it to define, manage and deploy features for both machine learning and large language model systems.

‘Completely ad hoc’

The state of feature management is “chaos” at many organizations, Khadder said. “It’s completely ad hoc. There’s no process at all,” he said, citing the example of one bank that had a library of 3,000 features “and they had no idea which ones were being used. A lot of them were defined by people who’d since left the organization, and everyone was afraid to turn them off because they didn’t know what they were and what was running through them.”

Featureform is an immutable platform, meaning that entries can’t be changed or altered after they are created. “We keep track of dependencies, so if you create a new one there’s still a pointer to the old one,” he said. “You don’t have to worry about all that complicated logic.” Because the platform tracks data lineage it can also detect and mask personally identifiable information.

Khadder said his company has seen a “huge explosion of interest since generative AI took off a year ago. “This space has matured as every cloud vendor has released a feature store,” he said. “We get a lot of inbound inquiries.”

The seed funding was led by GreatPoint Ventures LLC and Zetta Venture Capital LLC with participation from Tuesday Capital Management LLC and Alumni Ventures Group LLC. The company said the money will be used to “expand product growth and increase support for existing and new enterprise customers, in addition to its open-source community.”

Image: Featureform