Visa Ai Edge: How RAG-AS-A-Service and Deep Training Enhance Security and Acceleration of Data Extraction

A global giant for global payments Visa Operates in 200 countries and territories, all with their own unique, complex rules and regulations.

The team of his client services must understand these nuances when issues related to the policy-like “Do we have the right to process this type of payment in this country?” -But it is simply not possible to know all these answers in a human way.

This means that they usually had to track the relevant information manually – a comprehensive process that can take days depending on how accessible it is.

When a generative AI appears, Visa saw this as the perfect case of use by applying generation extracted generation (RAG) Not only to download information to 1000x faster, but also cite it back to its sources.

“First of all, these are more quality results,” Sam Hamilton, Visa’s SVP for data and AI, to Venturebeat. “This is also latency, isn’t it? They can deal with many more cases than they have been in a state before. “

This is just one way that a visa is used Like you To improve his operations-underlined by a deliberately built, multi-stage technological stack, while managing the risk and maintenance of fraud.

Secure Chatgpt: Visa Protected Models

November 30, 2022, the day Chatgpt has been introduced into the world will enter history as a major moment for AI.

Not long after, Hamilton noted, “Visa employees all asked,” Where is my chat? “” Can I use Chatgpt? ” “I can’t access Chatgpt.” “I want chat.”

However, as one of the largest digital payments suppliers in the world, VISA naturally had concerns about the sensitive data of its customers -more special that it remains secure, outside the public area and will not be used in the future Model trainingS

To respond to the search for employees while balancing these concerns, Visa introduces what he calls “Secure Chatgpt”, which sits behind a firewall and works internally at Microsoft Azure. The company can control the input and display by screening to prevent data losses (DLP) to ensure that sensitive data leave VISA systems.

“All hundreds of five data, everything is encrypted, everything is sure at rest, and also in transport,” Hamilton explained.

Despite the name, Secure Chatgpt is a multi -modern interface offering six different options: GPT (and its various iterations), Mistral, Anthropic’s Claude, Meta, Google’s Gemin Llama and IBM Granite. Hamilton described this as a model as a service or RAG-AS-A-Service.

“Think about it as a kind of layer in which we can provide an abstraction,” he said.

Instead of building their own vector databases, they can choose API, which best suits their specific use. For example, if they just need a little fine setting, they will usually choose a smaller open code model like Mistral; In contrast, if they are looking for more than a sophisticated model for reasoning, they can choose something like Openai O1 or O3.

In this way, people do not feel limited or seem to miss what is easily accessible in public space (which can lead to “shadow AI” or the use of disapproved models). Secure GPT is “nothing more than a shell on top of the model,” Hamilton explained. “Now they can choose the model they want from above.”

Beyond the safe Chatgpt, all visas developers are gaining access to GitHub Copilot to help their daily encoding and testing. Developers use Copilot and Plugins for various integrated development environments (IDE) to understand the code, improve the code and test units (defining this code is executed as intended), Hamilton noted.

“So code covering (identifying areas where there is no proper testing) is increasing significantly because we have this assistant,” he said.

RAG-AS-A-Service in action

One of the most powerful use cases for a secure chat is the processing of policy-specific policy issues.

“As you can imagine, as in 200 countries with different provisions, the documents can be thousands and thousands, hundreds of thousands,” said Hamilton. “This is really complicated. Do you have to nail this, right? And this must be a comprehensive search. ”

Not to mention that local policy changes over time, so Visa experts should be up -to -date.

Now with a sturdy rag based on reliable, up -to -date data, Visa’s AI not only quickly derives answers, but provides quotes and starting materials. “This tells you what you can do or you can’t do and say,” Here’s the document you want, I give an answer based on that, “Hamilton explained. “We have narrowed the answers with the knowledge we have embedded in the rag.”

Usually the comprehensive process will take “if not hours, days” to draw specific conclusions. “Now I can get it in five minutes, two minutes,” said Hamilton.

Visa birthday infrastructure with four layers of Visa

These opportunities are the result of Visa’s heavy investment in data infrastructure over the last 10 years: the financial giant has spent about $ 3 billion on its technological stack, according to Hamilton.

He describes that the stack as a “4 -layer birthday cake”: the foundation is a “canvas layer layer as a service,” data service “, AI ecosystem and machine learning (ML) and layers of data and products built on top.

Data-Platform-AS-AA-Service essentially serves as an operating system built on the data lake that aggregates “hundreds of fifth data,” Hamilton explained. The layer above, data as a service, serves as a kind of “data highway” with multiple lanes that run at different speeds to power hundreds of applications.

Third coat, AI/ml ecosystem is where Visa is constantly testing models to ensure that they perform as they should, and are not susceptible to bias and drift. Finally, the fourth layer is where VISA builds products for employees and customers.

Blocking a fraud by $ 40 billion

As a reliable payment provider, one of the main priorities of VISA is the prevention of fraud and AI plays an increased role here. Hamilton explained that the company had invested more than $ 10 billion to help reduce fraud and increase network security. Eventually it helped the company Block $ 40 billion when trying to cheat Only in 2024

For example, a new Visa Deep allowance tool provides an assessment of the risk of transactions to help manage payments on the card-not present (CNP) (for example, when users pay via web or mobile application, as is the daily practice for all of us). This is powered by a recurrent neural network (RNN) model based on fifths of contextual data. Similarly, real-time security protection (consider through digital portfolios or immediate payment systems)-AI deep-learned AI models that produce immediate risk results and automatically block bad transactions.

Hamilton explained that Visa uses a transformer-neuronic network-based model that learns context and meaning by tracking data in data-to improve these tools and quickly identify and thwart fraud. “We wanted to do this in accordance with transactions,” he said. “This means that we have less than a second, I have to say milliseconds, times of reaction.”

Synthetic data also provide value when fraud prevention: Hamilton’s team increases existing data with synthetic data on performing simulations around more fraud. “This helps us to learn what is happening now and what can happen in the short term and the long term, so we can simulate and train the model to capture the data,” he said.

He noted that the fraud is a race for weapons – and has a very low barrier to enter the threat participants. “We need to be a step ahead of this and predict and block them,” Hamilton stressed.