Discrepency.ai CEO Lisen Kaci on the Amiral Ventures Podcast

This May, our CEO Lisen Kaci sat down with the Amiral Ventures podcast for a chat about all things AI. Montreal-based Amiral Ventures is an early stage venture capital fund. The fund invests in tech startups with enterprise potential to drive strong, scalable growth at home. Without Canadian investors to take the lead early, many startups sell too soon, often to international investors. Amiral supports talented and passionate Canadian founders with capital, helping to create the next generation of future leaders. 

We sat down for a candid chat about AI engineering, document processing, and where we’re headed next. Here’s some of the highlights from our conversation. 

Overcoming Limitations of OCR

Founder Lisen Kaci built Discrepancy.ai out of firsthand experience – and a love of meeting customer needs. He started his career as an AI engineer at a document-processing company in Toronto. After witnessing the inconsistencies created out of traditional systems powered by optical character recognition (OCR), he could see where things needed to be improved. 

Pulling financial data out of documents is difficult with OCR. Even though financial information should theoretically be an easy task, something like a bar graph is easy for humans but difficult for AI. Lisen spoke about companies trying to preserve legacy OCR capabilities while adding on pieces of more modern AI – a process, he says, that just doesn’t fit. Discrepancy was built from scratch based on current capabilities.

Traditional tools, like OCR, use probabilities to predict labels and relationships. They don’t understand the data, and thus require more manual oversight. Current technologies, like Natural Language Processing, can process language in context and understand it. While manual review is often beneficial, it’s not the best option for repetitive or time consuming tasks. NLP helps take manual labor out of a repetitive process without sacrificing accuracy. 

“Let's say a customer wants to do income verification,” Lisen explains. “That's a really big use case for us. Whether it's someone screening like rental screening, mortgage screening, loan applications, there's so many different types of use cases for this one task of income verification. Right now, most companies are like 80% manual review of these documents.”

“For OCR, every single type of bank statement requires its own model – you have to fine-tune, so one for RBC, one for TD, one for all of these credit unions…and then the same with pay stubs. This is really a limiting factor with how the current OCR technology works and really where our use case shines. We can save 80% on labour costs for reviewing these documents and you don't need to have a separate AI model for each.”

Asimov’s Hand: AI Tool Chaining in Modern AI

In order to keep things modern, Lisen replaced legacy OCR tools with modern AI, including NLP. However, that’s not the only ‘AI facelift’ that Discrepancy has given to the system. Rather than relying on one single purpose model for the platform, Discrepancy is built with multiple AI models chained together, one after the other. 

“There’s an internal ranking for what’s better for each task, like subtraction or extraction, and we’ve built a system that picks the best one and helps all of these models work together,” Lisen explained on the podcast. He brought up Isaac Asimov’s analogy of the ‘hand and the fingers’ from I, Robot. Using one main AI model built to distribute the tasks and workload to many smaller ones gives you a faster, more flexible, system capable of harder tasks. Not to mention, the ‘little robots’ help catch each other’s hallucination, making for a much more refined final result. 

What’s Next? The Future and Scale

From financial services to proptech and insurance, Lisen says his vision is simple: to help organizations scale document heavy workflows without scaling headcount. By automating more of the tasks and tailoring each solution to each user, he told Amiral Ventures’ host that “at the end of the day everything's an SQL wrapper. Everything's a wrapper of something, so it's just about how can I solve my customer’s problems in a really good way.” 

We’re proud to be building in Canada and grateful for the time to talk tech with the Aminal Team! The full interview is available on Aminal’s YouTube.