Person
1 episodes · 19 quoted claims
Sumedh Chaudhary says he has over twenty years in the industry and that the principle of garbage in, garbage out has held consistently.
“I've been in the industry for over twenty years and the one consistent thing I've heard over the years is any garbage you you give into an enterprise system, you'll get garbage out.”
Chaudhary says enterprises have recently had success using generative AI models to analyze unstructured data.
“Uh, using some of the generative AI models to be able to analyze and put solutions for unstructured data.”
Chaudhary says documents combine images, tabular data and page breaks, which add complexity beyond ordinary unstructured data.
“you've got images in it, you've got tabular data in it. And they are something we call as page breaks, right?”
Chaudhary describes an OCR agent whose sole focus is recognising data and extracting a machine-usable digital footprint.
“you have an optical character recognition or an O C R agent.”
Chaudhary says vector agents take embeddings from extracted data and store them in a vector database.
“Then you have vector agents that take the embeddings out of that extracted data and be able to store in a vector database”
Chaudhary describes a splitter agent that preserves context across consecutive pages of a document.
“So you can have something called, we call it as a splitter agent, so you understand the context between page one, page two, page three.”
Chaudhary says he has recently seen many users upload documents running to thousands of pages.
“I've seen a lot of folks recently using documents that are thousands of pages.”
Chaudhary describes a matching agent that links references across multiple uploaded documents.
“is what I call as a matching agent. And this is an agent that when you're uploading multiple documents into the system, is able to give references”
Chaudhary says proofs of concept succeed when scope is small but may fail in production.
“where you have a very successful proof of concept when the problem and the scope of the solution was small.”
Chaudhary recommends tracking performance and error-rate targets week by week.
“you should have target numbers for your performance and error rates that you're tracking week by week.”
Chaudhary suggests promising a CFO that an AI solution can match legacy benchmarks within three to six months.
“we can achieve the same benchmarks as a legacy solution in a period of three months or six months.”
Chaudhary says a successful agentic workflow is no longer limited by the number of human users who can handle it.
“You do not have to, you're not constrained by the number of, uh, users who are able to handle that workflow anymore.”
Chaudhary says many automation use cases for data access can be done before adopting agentic workflows, sometimes without an AI agent at all.
“Before we jump into the agentic workflows, there's a lot of automation use cases that can be employed in terms of how you access the data”
Chaudhary uses a sedan versus pickup truck analogy to argue that different enterprise tasks need different AI models.
“A small sedan car probably does not need the same engine that you need for a heavy-duty pickup truck, right?”
Chaudhary says enterprises can access thousands of models, though not all are available in every cloud environment.
“an enterprise can have access to thousands of models, and some of those models may or may not be available to you in your cloud environment.”
Chaudhary compares orchestrating AI agents to a concert master leading an orchestra of talented musicians.
“you've got this concert master who is trying to run this beautiful orchestra.”
Chaudhary advises first mapping all data sources and reaching them through a common virtualized layer.
“you spend a lot of initial time making sure you're mapping all your data sources and you have access to that in a in a common tool, work virtualized layer”
Chaudhary argues leaders should not rush into a use case merely to claim an early win.
“And I would not expect leaders to rush into a use case just because they wanna check a box and get an early win.”
Chaudhary says organisations whose AI error rates miss compliance standards may revert to legacy manual systems.
“you might go back to your old systems, the legacy systems, where you were doing the works a little bit more manually.”