Intellect Design Arena scales trusted ESG AI

Beyond the model: governing enterprise knowledge at scale

In the highly regulated world of global finance, moving Artificial Intelligence from experimental prototype to production is notoriously difficult. For Intellect Design Arena, a product and platform company serving some of the world's largest banks, insurance firms, and financial services organizations, navigating this complexity is fundamental to its work. Operating across global geographies and the banking lifecycle, from consumer to corporate, means Intellect understands the regulatory and governance requirements its clients face every day.

“It means for us we need to worry about regulation a lot and we need to worry about governance a lot,” says Deepak Dastrala, CEO, Purple Fabric.

The company recognized a critical industry blind spot: while organizations often focus heavily on improving their underlying AI models, operationalizing AI in highly regulated sectors requires much more than model performance. At enterprise scale, data governance, architecture, security, and reliability all need to be considered from the outset. The challenge is not simply accessing enterprise knowledge, but ensuring it can be used safely, securely, and consistently.

To address this challenge, Intellect developed Purple Fabric, its dedicated enterprise AI platform. As part of its broader technology ecosystem, Intellect partnered with MongoDB to explore an architecture capable of supporting AI at scale in a highly regulated environment. A major proving ground for this approach was a real-world use case involving complex Environmental, Social, and Governance (ESG) data for a major sovereign wealth fund.

The Challenge of Unstructured ESG Data

The core problem the wealth fund faced was that its critical ESG information extends well beyond standard climate metrics to include complex topics such as oceans, taxes, and child labor. Much of this information is not neatly structured, but instead distributed across PDFs, images, public web documents, and other sources.

This fragmentation limited the fund's ability to evaluate its global portfolio comprehensively.

“Before we brought this platform together, they were able to do it only for a fraction; 2% of the companies they have invested in, on a very selective basis,” explains Dastrala.

Making informed decisions about whether to invest or divest billions of dollars requires broad and reliable data coverage. The fund therefore needed an approach capable of processing diverse enterprise information while maintaining data lineage, relevance, and governance.

Rethinking the retrieval architecture

To address the challenge, Intellect designed the retrieval architecture within Purple Fabric around a combination of technologies and approaches, including MongoDB Atlas. The focus was not simply on selecting a database, but on creating an architecture that could support reliable retrieval, efficient processing, and enterprise-scale AI applications.

“For me the biggest reason for this choice is not just cost and performance, it is also the stability,” explains Dastrala. “Being completely available within MongoDB Atlas has given me the flexibility that I’m looking for.”

A key architectural decision was to separate data gathering from the application and reasoning layers. Rather than relying solely on semantic or vector search, the team adopted a hybrid approach that combined AI-driven retrieval with traditional lexical search. Metadata filters could also be applied early in the process to narrow down large volumes of information before more intensive AI processing took place.

As Dastrala advises, “For anyone who wants to productionize AI, focus more on architecture, focus more on engineering. Models will continue to get better, but that will not help you to move to production.”

The architecture also enabled the team to evaluate retrieval, relevance, and answer accuracy as distinct stages. This separation provided greater visibility into where issues occurred and helped the team build a more reliable approach to answering complex questions around corporate environmental and social practices.

In testing, vector-only search could identify broader concepts but sometimes missed specific company details and dates. The hybrid approach was designed to address both dimensions—understanding the broader context while also retrieving precise information.

Improving latency and accuracy

The resulting architecture brought multiple retrieval capabilities together within a unified pipeline, including metadata filtering, vector search, lexical search, and reranking. This reduced the need to coordinate multiple separate components across the retrieval workflow.

The results demonstrated the value of the approach. An initial vector-only search achieved an accuracy rate of 71%. With hybrid search and Reciprocal Rank Fusion (RRF), accuracy increased by 17%. P95 latency—the time by which 95% of requests had returned—was also reduced by approximately half, improving the responsiveness and stability of the system.

The solution was taken from implementation to production in approximately six to eight weeks.

“It is not just a cost problem or a performance problem, it is actually a stability problem,” says Dastrala, emphasizing the importance of maintaining tight, relevant context rather than indiscriminately providing more information to an AI model.

“More context is great,” he adds. “But it’s about the right context.”

Designing for trust and governance

In highly regulated financial environments, precision and accountability are critical. While the platform is designed to operate with a high degree of autonomy, Intellect has deliberately incorporated human oversight for cases where the system's confidence or available information is insufficient.

For the remaining 9% of highly ambiguous queries, the platform triggers a human-in-the-loop workflow, allowing domain experts to review cases where data may be incomplete or the answer requires additional judgment.

Governance is also built into the process from the point of data ingestion. Rather than treating governance as a downstream control, Intellect embeds policies and controls within the pipeline before information is exposed to the LLM.

“Governance cannot be an afterthought,” says Dastrala. “You have to enforce it in the pipeline so that it’s taken care of before you expose anything to the LLM.”

“Knowing where your AI is unreliable is actually part of making it trustworthy,” he adds.

The experience demonstrates a broader principle for enterprises looking to operationalize AI: successful deployment depends not only on the model, but on the architecture, data, retrieval strategy, governance, and engineering practices surrounding it. Through Purple Fabric and its broader technology ecosystem, including its collaboration with MongoDB, Intellect Design Arena is building an approach designed to help financial institutions innovate responsibly while meeting the demands of enterprise scale.