Data science teams are stymied by disorganization at their companies, impacting efforts to deploy timely AI and analytics projects. In a recent survey of “data executives” at US-based companies, 44% said that they’ve not hired enough, were too siloed-off to be effective, and haven’t been given clear roles. Respondents said that they were most concerned about the impact of a revenue loss or hit to brand reputation stemming from failing AI systems and a trend toward splashy investments with short-term payoffs.
These are organizational challenges. But Piero Molino, the cofounder of AI development platform Predibasesays that inadequate tooling often exacerbates them.
“The major challenges we see today in the industry are that machine learning projects tend to have elongated time-to-value and very low access across an organization. As a result, most machine learning tasks in an organization are bottlenecked on an oversubscribed centralized data science team,” Molino told TechCrunch via email. “Given these challenges, organizations today need to choose between two flawed approaches when it comes to developing machine learning. They can build-their-own systems from data-to-deployment using low level APIs that give them the flexibility machine learning tasks typically require at the cost of complexity. Or they can choose to use a blackbox off-the-shelf ‘AutoML’ solution that simplifies their problem at the expense of flexibility and control.”
Indeed, while worldwide spending on AI technologies was estimated at $35.8 billion in 2019, nearly 80% of companies have seen their AI projects stall as a result of issues with data quality and a lack of confidence in AI systems, according to an Alegion report. Being an entrepreneur (and a salesperson), Molino asserts that his product, Predibase, is a solution to this — or at least a step toward one.
Predibase, which today emerged from stealth with $16.25 million in Series A funding led by Graylock with participation from the Factory and angel investors, allows a user to specify an AI system as a file that tells the platform what the user wants (eg, recognizing objects in an image) and figures out a way to fill that need. Molino describes it as a “declarative” approach to AI development, borrowing a term from computer science that refers to code written to describe what a developer wishes to accomplish.
“Machine learning projects today usually take six months to a year at most organizations we’ve worked with. We want to drastically reduce that [by bringing] a low-code but high-ceiling machine learning tool to organizations” Molino continued. “Typically, most companies are bottlenecked by data science resources, meaning product and analyst teams are blocked by a scarce and expensive resource. With Predibase, we’ve seen engineers and analysts build and operationalize models directly.”
Predibase is built on top of open source technologies including Horovod, a framework for AI model training, and Ludwig, a suite of machine learning tools. Both were originally developed at Uber, which several years ago transitioned governance of the projects to the Linux Foundation.
Molino, who joined Uber by way of the company’s acquisition of startup Geometric Intelligence, helped to create Ludwig in 2019. Predibase’s other cofounder, Travis Addair, was the lead maintainer for Horovod while working as a senior software engineer at Uber.
To launch Predibase, Molino and Addair teamed up with former Google Cloud AI product manager Devvret Rishi and Stanford computer science professor Chris Ré, one of the cofounders of Lattice.io, a data mining and machine learning company that Apple purchased in 2017.
Predibase is designed to enable developers to define AI pipelines in just a few lines of code while scaling up to petabytes of data across thousands of machines. As Molino explains it, using the platform, a user can create a text-analyzing AI system in six lines of code that specifies the input and output data. If they want to iterate and customize that system, Predibase lets them add parameters in the configuration file that affords a more granular level of control.
Predibase integrates with data sources including Snowflake, Google BigQuery, and Amazon S3 for model training. Users can train models through the platform or programmatically, depending on the use case, and then host and serve or deploy those models into local production environments.
Apart from lowering time-to-value, Predibase allows users to work with different modalities of data using the same toolset. With Predibase, we’ve seen user train models on images for classification, text data like emails for triage, tabular data for detection and regression tasks, and even audio datasets that would’ve required heavy in-house sophistication without the native capabilities in the platform,” Molino said. “For many working in this space, Predibase provides a net new capability when tackling use cases on unstructured data.”
Broadly speaking, no-code development platforms are on the riseand a number of startups compete directly with Predibase, including AI orchestration startup Union.ai and low-code data engineering platform Prophecy (not to mention SageMaker and Vertex AI). But Molino’s view is that while rivals satisfy the demand in the enterprise for simple solutions, they do so at the cost of flexibility, leading customers to “hit a ceiling and churn out.”
“[L]ike infrastructure as code simplified IT, our platform allows users to focus on the ‘what’ of their models rather than the ‘how,’ allowing them to break free of the usual limits of low-code systems using an extensible configuration … We provide model explainability out-of-the box so users can understand which features are driving predictions,” he said. “[Our platform] has been used at Fortune 500 companies like a leading US tech company, a large national bank and large US healthcare company.”
The pitch sufficiently impressed angels like Kaggle CEO Anthony Goldbloom and former Intel AI COO Remi El-Ouazzane, both of whom invested. Other notable backers include Kaggle CTO Ben Hamner and Zoubin Ghahramani, a professor of information engineering at Cambridge and senior research scientist at Google Brain.
Molino says that the fresh capital from the Series A will be used to take Predibase’s beta product to a wider market — it’s currently invite-only. It’ll also be put toward growing Predibase’s team of machine learning engineers and building out a go-to-market organization, expanding the company’s 21-person team.