Cloudera aims for data architecture that supports enterprise AI – News

Designing a data architecture that can truly support artificial intelligence models is proving to be one of the toughest challenges for enterprises.

Cloudera Inc. is addressing the issue by taking a flexible approach, allowing AI to operate where the data already resides — whether in on-premises systems, sovereign frameworks, or the cloud. The goal is to give enterprises both regulatory compliance and a seamless user experience as AI adoption accelerates.

“The models need to be located where the data is,” said Manasi Vartak (pictured), chief AI architect at Cloudera. “If you have a fraud model for Europe, run it within Europe. Bring your models where your data resides so you can both stay within regulation and to give a better user experience. If you start shipping data across the world to get an LLM prediction, you’ve lost the game.”

Vartak spoke with theCUBE’s Dave Vellante at Cloudera EVOLVE25, during an exclusive broadcast on theCUBE, News Media’s livestreaming studio. They discussed the growing impact of agentic and generative AI and how Cloudera is charting its own path. (* Disclosure below.)

Data architecture that supports an AI workforce

Vartak started her journey as the founder of machine learning infrastructure startup Verta, which was recently acquired by Cloudera. She has translated that background in machine learning and predictive AI to Cloudera’s platform, taking on the mammoth task of managing AI models and agents in the enterprise environment.

The current strategy is combining different types of AI within a flexible data architecture to get the best results for the given problem, according to Vartak.

“Predictive AI is also here to stay,” she said. “If you’re looking at a fraud detection use case, you might use generative AI and embedding models that came with generative AI in order to produce features that then go into your predictive AI and then tell you is this fraudulent or not.”

The industry is expecting the rise of agent workforces working alongside human ones. Before that can become a reality, leaders such as Vartak need to determine access and controls for AI. Many of Cloudera’s customers, concerned about privacy, have leaned toward using open-source models with proprietary data and hardware.

“If you think about the digital workforce, they have a role,” Vartak said, “And depending on the role, they have access to certain data, they cannot have access to everything because then you’ll have, say, a writer, a summarizer agent who has access to your financials. That would be terrible.”

Cloudera’s goal has been to create a single user experience, whether the customer is running AI at the edge, in the cloud or within a sovereign framework. The company wants to keep the platform open and extensible, so customers can bring in any new AI tools.

The advent of business-specific agents models will continue to bring up issues around hallucinations, but Vartak suggests that gen AI’s imperfect results can sometimes be part of its charm.

“Hallucinations are an artifact of how these models are trained,” she said. “OpenAI just tested out different ways to train the model to reduce hallucinations, but this risk is if you reduce hallucinations, it also reduces the creativity of the results. There’s ways to circumvent that and put in guardrails or extra checks to mitigate hallucinations, but I think we need to accept that they’re going to be a part of gen AI outputs.”

Here’s the complete video interview, part of News’s and theCUBE’s coverage of Cloudera EVOLVE25:

(* Disclosure: TheCUBE is a paid media partner for Cloudera EVOLVE25. Neither Cloudera Inc., the sponsor of theCUBE’s event coverage, nor other sponsors have editorial control over content on theCUBE or News.)

Photo: News