Lisen Kaci, Discrepancy AI CEO, was featured on the Industry Live Podcast to discuss Digitizing Data for AI

The Industry.live podcast explores how leaders use emerging tools like AI to solve real problems that businesses face. Hosted by Soumik Roy, the podcast shares insights from strategy, governance, and technology innovators. Lisen Kaci sat down with the Industry.live podcast for a chat about Discrepancy, startups, and all things OCR. 

Lisen talks about building Discrepancy amid the first wave of omni-channel LLMs, his journey from AI engineer to startup founder, and shared more about why businesses are stuck using outdated OCR. 

How AI Helps Fix Document Processing and Manual Workload

There has been much said about AI and document processing – but did you know that AI document processing actually lags behind in most enterprise organizations? Lisen talks about how his work in AI, as both a lead AI engineer and running another AI startup pre-GPT3, exposed him to some inefficiencies in the industry. He noticed that while cutting edge, omni-channel LLMs existed on the market, enterprise organizations weren’t using them. Complex document processing and workflows were still being done by hand, so he decided to build what was missing. 

The Unstructured Data Problem 

The current state of document processing for enterprise companies is generally optical character recognition (OCR). The unstructured nature of documents (such as medical documents or financial reports) makes it difficult for OCR tools to understand; technologies like OCR don’t understand the meaning of the data that they pull. OCR uses probabilities to predict labels and relationships within data. A circle on a page, for example, might have a 72% chance of being a letter O, an 18% chance of being 0, a 6% chance of being Q and a 4% chance of being part of a smiley face. 

More current technologies, like Natural Language Processing, can process language in context. The same circle might be interpreted with 99% certainty in “my account number is 750” or “his name is spelled T-O-M” because of the words and letters that surround it. 

To keep things modern, Lisen replaced OCR with more modern NLP systems in a system that chains multiple models together – rather than relying on one, single purpose, AI model, Discrepancy is built with one powerful model that distributes tasks and workloads to smaller AI agents, capable of a smarter, more flexible system capable of hard tasks. These smaller models help catch each other’s mistakes, leading to a more refined result. For example, data found in documents, images, or PDFs, gets passed into structured JSON data.  

“We fine tuned our own models for some specific use cases,” Lisen says, “And we have our own internal ranking about what model is good at what specific task. When a document comes in, we run it through these multiple sequential models. One to extract the data, one to standardize it, one to gather insights from it, and another to look at fraud.”

At the end of the day, Lisen says modern AI is alchemy. “You could figure out a way to combine things and get gold. In terms of advancements in how AI models work and what they can do, we are scratching the surface.”

For business users facing rising workloads and a plethora of new AI tools, this episode is a must-listen! Thank you to Soumik Roy and the Industry.live team for giving us the opportunity to talk AI, alchemy, and everything in between.