Broadridge Financial Solutions Inc. has been granted a U.S. patent covering its proprietary methods for orchestrating machine learning agents via large language models (LLMs), a technology central to its BondGPT and BondGPT+ platforms. The tools are designed to support corporate bond trading and analytics through natural language processing and AI-driven data orchestration.
Initially introduced in June 2023 through Broadridge subsidiary LTX, BondGPT was followed by the enterprise-grade BondGPT+ version in October. Both platforms leverage OpenAI GPT models and Broadridge’s orchestration technology to respond to user queries by pulling and analysing data from multiple datasets and analytical models in real time. BondGPT+ integrates client-specific data and models, enabling buy- and sell-side users to access personalised insights and pre-trade analytics more efficiently.
“When we launched BondGPT, we set out to solve a core challenge our clients faced. How to unify the vast and varied data available to bond traders, from proprietary LTX data like Liquidity Cloud to widely used sources such as TRACE and pricing data,” explains Jim Kwiatkowski, CEO of LTX, in conversation with TradingTech Insight. “The goal was to let users ask simple, natural language questions, much like early use cases of ChatGPT, and get meaningful, accurate answers that draw from all those datasets. To achieve that, we developed a system of AI agents, each with a specific role. Some to understand and parse the query, especially when it involved identifiers like tickers or CUSIPs, and others to retrieve and process the relevant data or models. We quickly realised that out-of-the-box LLMs weren’t enough; we had to build a tailored architecture to meet the specific needs of bond trading. Compliance was also a critical layer. As a broker-dealer, it wasn’t just about answering the question, we had to ensure every response aligned with FINRA regulations. Each of these needs effectively became an agent in our system, long before the term ‘AI agent’ became as mainstream as it is today.”
The newly patented technology under U.S. Patent No. 12,061,970 introduces a number of advanced features, including a “Show your work” function to explain how outputs are generated.
“Explainability is a central element of this patent,” says Kwiatkowski. “While ‘Explainable AI’ has been a familiar concept for years, we approached it from a trader’s perspective. Imagine you’re on a trading floor and ask an assistant: how many bonds traded below issue price last week, are rated triple-B or better, and come from issuers with EBITDA growth over the last three years? The assistant would need to gather data from TRACE, financial filings, ratings, and pricing sources, do the analysis, and return a few results. Now, if you’ve worked with that assistant for a decade, you’d probably trust the answer. But if the assistant was new, you’d want to know how they got there. You’d say: ‘Show me the work.’ That’s exactly how we thought about AI—it’s like the new assistant. So we built a feature where you can click a button and see a step-by-step breakdown of what the system did: where it got the data, what models it used, how it combined the inputs, and how it arrived at the final output. Traders may not check it every time, but especially early on, it builds trust. That level of transparency, delivered in a conversational way, is something we believe is genuinely unique.”
The patent also includes a multi-agent adversarial system to enhance accuracy, and an AI-powered compliance verification tool that supports custom enterprise risk frameworks.
“One of the agents we developed is an adversarial agent, designed to challenge the outputs of other agents before an answer is delivered. This plays a key role in ensuring both accuracy and compliance,” states Kwiatkowski. “As a broker-dealer, we operate under different regulatory obligations depending on the context. To address this, we built an AI-driven adversarial compliance agent that can be configured to align with both LTX’s internal rules and those of our clients. It adds a critical layer of oversight, helping ensure that responses not only make sense, but also meet the appropriate regulatory standards.”
Additionally, the system uses user role profiles to tailor data access and security. “An analyst and a trader may get very different answers to a similar question because of their roles and the types of information they’re seeking,” says Kwiatkowski. “Similarly, a wealth manager querying one client account will receive information they’re entitled to, but if they ask about another account without access rights, the system will restrict the response. We’re already used to data systems that allocate access based on user profiles, so we built that principle into the AI. The agents are aware of who the user is, not just for security, but for relevance. For example, if someone is a high-yield trader, the system won’t return results about investment-grade bonds. It focuses on what matters to that user. That concept of profile-driven entitlements – both for access control and contextual accuracy – is also a core part of the patented architecture.”
The patent adds to Broadridge’s growing intellectual property portfolio supporting LTX’s fixed income trading solutions, which already includes innovations such as bond similarity scoring, dealer selection algorithms, liquidity aggregation, and the RFQ+ trading protocol.
Subscribe to our newsletter
