Person
1 episodes · 19 quoted claims
Kapireddy's group at Truist is responsible for AI strategy and policy, data science platforms, and developing advanced machine learning and generative AI applications.
“The third bucket is really developing advanced machine learning and gen AI capabilities, including the applications themselves.”
Truist's traditional AI already includes fraud detection and customer segmentation models.
“we have number of AI models that are already in place from fraud detection models, we have some customer segmentation models”
Kapireddy says Truist's generative AI approach includes using services such as Azure OpenAI and Amazon Bedrock, which has Anthropic models behind it.
“using the services, APIs, Azure Open AI is as a service, for example, the bedrock as a service, uh, with the anthropic behind the scene.”
Kapireddy says the 2017 paper 'Attention Is All You Need', which introduced transformer architectures, was when the field changed.
“as a matter of fact, 2017, I think, is when it all changed with attention is all you need and the transformer architectures.”
Kapireddy describes a credit-card underwriting model as built to produce a score that aids approval decisions and requiring very high explainability.
“That model was specifically built to either give a score to aid the decision for credit card acceptance or rejections.”
Kapireddy says a single generative model can summarize text, extract entities and assess sentiment, tasks that previously needed separate models.
“Maybe you want to summarize the text, maybe you wanted to extract entities out of the text, maybe you want to get the sentiment of the text.”
Kapireddy says NIST released a generative AI profile to supplement its AI Risk Management Framework after generative AI went live.
“NIST had to release a generative AI profile to aid the NIST AI risk management framework.”
Kapireddy says Truist uses a seeding mechanism so a generative model gives the same output for the same question.
“Fortunately, we've been able to implement the seeding mechanism. So it actually at least gives the same output no matter how many times you ask the same question.”
Truist's AI lifecycle has seven stages, beginning with ideation and risk tiering and continuing through development, testing, independent validation, implementation and ongoing monitoring.
“Then we actually begin to the development, testing, independent validation, and implementation and ongoing monitoring. This is seven stages of life cycle.”
Kapireddy says no financial-services use case lets generative AI make customer-affecting decisions without a human in the loop.
“I don't think there is any use case that is actually a customer facing affecting the decisions that we would make without human in the loop as an industry.”
Kapireddy says Truist tells users they are interacting with AI and warns that it could hallucinate.
“we will make it clear that, hey, you're interacting with an application that is AI, and we also warn them that it could hallucinate, and so people can take the appropriate actions.”
Kapireddy says tools like Microsoft Copilot are used poorly without training in prompt writing and understanding the system's limits.
“If you do not understand how to write a prompt to get an answer, and if you don't understand on how the system would actually behave”
Kapireddy says generative AI is a favored tool for attackers, making fake voices and fake invoices more sophisticated.
“Gen AI is definitely most favorable tool for folks who really want to take advantage of it to increase their attacks, right?”
Kapireddy notes generative adversarial networks make it increasingly hard to tell real content from fake.
“I'm sure you're all very familiar with the GAN models, the generative adversarial networks model.”
Kapireddy says Truist would use a deterministic API call, not a generative model, to answer a customer's checking balance.
“So we're going to have a deterministic API call that agent has to make to get a specific answer for that.”
Kapireddy describes retrieval-augmented generation (RAG) patterns, where relevant information is retrieved and added to the prompt before the language model is called.
“you obviously contextualize it by extracting information relevant to the prompt, and then you send it to LLM, including making API calls to get more deterministic answers”
Kapireddy says Truist builds its own guardrails instead of relying only on those provided by service vendors.
“companies like Truist, we build our own guardrails to make sure that the output is thorough and ensure that this output can be trusted and relied upon by the end user.”
Kapireddy says he once weighed film-direction training in India against a US master's degree and chose the US because of database management.
“I think when I was in India, I was contemplating to either to come to United States to do my master's”
The hosts say MIT Sloan Management Review and Boston Consulting Group have researched and published on AI together since 2017.
“Together, MITSMR and BCG have been researching and publishing on AI since 2017, interviewing hundreds of practitioners and surveying thousands of companies”