Racing down a mountain on a bike while simultaneously racing to deliver a groundbreaking GenAI-powered application is a hazardous combination. Joseph Lo, Head of Enterprise Platforms at Broadridge, was juggling this tricky mix as a deadline loomed in the development of BondGPT, a cutting-edge GenAI-powered application from Broadridge that answers complex bond-related questions in seconds.

“We had a decision to make. Do we have a viable product or not? We had to solve many technical problems quickly to make that call. There was one weekend in May 2023 when it all came to a crunch – and I also happened to be in a mountain biking race. We were trying to do things with GPT 3.5, GPT 4 that had never been done before,” explained Joseph. 

“I remember that weekend well,” said Jim Kwiatkowski, CEO of LTX, a Broadridge company. “We had been fervently working around the clock to advance this major project. In the thick of a product development brainstorm Teams chat with Joseph, he warned that his responses might be delayed because he was competing in a mountain bike race! I was a little alarmed, as you can imagine.”

Joseph Lo, Head of Enterprise Platforms at Broadridge, races down the mountainside.

Broadridge was working on a solution to a huge challenge in the world of corporate bond trading.

“We have incredibly rich and unique bond data on our LTX platform,” said Fitim Kryeziu, VP of Data Insights at Broadridge. “We had a patented model called ‘Similar Bonds’ but it had a complex user interface. Then one of our advisory board members asked us to find a simpler way to use our data and models to assist in his pre-trade workflow. He believed that if we could make the data and models easily accessible in a simple interface within the trading desktop, this would solve real-world challenges faced by corporate bond traders in their daily workflows - and serve as a differentiator for LTX.”

Traders’ screens are incredibly congested, leaving little space for new tools. The LTX/Broadridge team initially attempted to solve the data access challenge by creating a new dashboard. Still, it proved a challenge as every trader wanted the data and information showcased differently, requiring many rows, columns, menus, and dropdowns, creating more complexity rather than minimizing it.

“Synthesizing trading data into something useful for trading decisions usually involves analysts, trader research teams, and hours of number crunching, revisions, many subsets of data, and more,” explained Jim. “Our rapidly assembled team wanted to return that kind of analysis within seconds and present it in a very small section of a screen. We also wanted it ready to demo in early June at the biggest event for fixed income leaders.”

ChatGPT makes an entrance

“It was right around this time that ChatGPT was released,” said Jim. “I was playing around with it and challenged the team: Can we create a tiny little text box and give every user the ability just to type an English language question and get the precise answer to their varied questions from any combination of our data and models?”

Joseph had experience with ChatGPT and other AI coding techniques. Fitim was an expert in fixed income data, both third-party data and Broadridge’s. Jim stepped up as the experienced voice of many different types of data users, like traders and portfolio managers. “I thought about the different types of questions different personas would want to ask,” Jim said. The team started work in March 2023, and, by the end of April, they knew they were onto something good.

The creativity problem

“The immediate problem was that ChatGPT had something special about it that was incredibly unhelpful for financial markets: creativity,” said Fitim. “It can write a poem or provide a speech in the voice of Abraham Lincoln, but there’s just no room for creative expression in trading. You must provide accurate answers 100% of the time.”

“We needed BondGPT to do bond math,” said Joseph. “It had to be accurate, and we didn't want it to get its data from the Internet. We used the OpenAI technology to determine what the bond trader, wealth manager, and portfolio manager were looking for. Then, we got the data only from our verified, curated databases.”

“We needed to improve on what ChatGPT did,” said Jim. “GPT understands the English language but doesn’t understand ‘bond speak’, the language spoken on a trading floor. We collectively needed to teach it that. I remember Joseph working through ticker symbols. How does it know when you say Apple, that Apple is AAPL? Or if you give it a ticker symbol, that was a ticker symbol and not a mistyped word? We tried to emulate the conversation as if you were asking a colleague the question.”

A question of trust

The team also faced a trust issue. Why would a trader trust the results from a brand-new natural language function that delivers in 20 seconds something that would usually take 20 minutes or more?

“If someone you had just met gave you an answer to an important question, you might ask them how they arrived at that answer,” said James. “It’s a polite way of saying ‘Sorry, but I don’t 100% trust you on that.’ So, we built in a ‘I don't 100% trust you’ button into the product, and it's called ‘Show Your Work’. With a click, it'll say, ‘I went to these data sources; I did this calculation; I presented the answer to you this way and that way.’ It provides complete transparency and an audit trail of its actions. Even ChatGPT doesn't do that.”

Nope. Too slow.

“When we were ready with our first prototype of BondGPT, after more than 500 iterations in just a few weeks, I proudly showed the mobile app to my daughter, a medical school student who knows very little about fixed income,” said Jim. “At dinner, I said, ‘Check out this cool thing we're working on.’ About 30 seconds after I handed her the phone, and before it answered the question, she returned it to me and said, ‘This is taking too long. Just tell me what it does.’ That was the moment I knew I had to go back and tell the guys that it needed to work within 30 seconds. Even though we'll save someone minutes or hours, if it can't respond in 30 seconds, the human attention span seems to hit this limit. And that’s when we set ourselves a new goal.”

“I'm happy to say that most of our requests are now answered in less than 10 seconds. And we're super proud of that because, again, we had to invent a lot of things to optimize that stack,” said Joseph.

The big launch with burgers and bourbon

“Jim had this very ambitious idea that we’ll launch on June 6 at the Fixed Income Leaders’ Summit, which is the largest event of the year for corporate bond traders. He wanted every person at the conference to have access to BondGPT.  We would be the buzz of the conference and everyone would be talking about LTX. That was one hell of a race to get the product to the point where it was ready. Whatever we said it was going to do, it needed to do well and fully,” said Fitim.

“It was one of the best moments of my career, for sure,” said Joseph. “We hosted a burgers and bourbon evening because the conference was in Nashville. The room was packed to the gills, and we had BondGPT on a big screen for demos. Attendees shouted out, ‘Try this question, try that question.’ They were so thrilled by how transformative it could be. That was an awesome moment.”

“Taking part in the process of ideating and developing the product from beginning to end was really unique as a marketer,” said Caitlin Mulkeen, Senior Director of marketing at Broadridge. “Though much time was spent covering all our bases for the marketing launch, from the big conference to our waitlist landing page to brochureware to press and media, we had no idea that the launch would resonate the way it did. Witnessing the palpable enthusiasm at the launch event was one of the most rewarding moments of my career.”