How We Built an AI Analyst That Actually Cites Its Sources (Not Just Another ChatGPT Wrapper)
AI has been a trending topic for the past few years. As a result, many individuals and companies keep rolling out AI products and tools designed to solve one or more problems across industries.
The finance industry has had its own wave of AI tools. However, there is a particular kind of AI product that we have become increasingly skeptical of: ChatGPT wrappers. These are apps that build custom solutions for a specific audience based on the model of platforms like ChatGPT or Claude.
In finance, ChatGPT wrappers feed financial data into LLMs using APIs, the most popular of which is a stock API.
However, we have discovered that while LLMs are good at synthesizing information, there is always the question of whether the information provided is relevant, current, and traceable.
Therefore, when we built Xurve AI Analyst, we were not creating another finance-focused GPT wrapper. Rather, we sought to build a comprehensive terminal with a rigorous context architecture that solves the problems of standard ChatGPT wrappers.
How did we do this?
Institutional-grade market data and analysis
There are three layers to the Xurve AI Analyst product.
At the bottom is the underlying market and financial data. This includes fundamental, technical, and sentiment data that are relevant to financial assets like stocks, currencies, commodities, and cryptocurrencies, among others.
The second layer is market news and market research. Market news covers the most relevant and recent conversations around a given asset, while market research contains a detailed overview of market sentiment, risk exposure, macroeconomic and industry impacts, and other potential price drivers.
The third layer focuses on extracting relevant entities, identifying relationships, and prioritizing information from the other two layers.
Real-time data integration
Standard LLMs do not provide real-time data, which often limits their usefulness for relevant analysis.
With Xurve, we sought to create an AI conversational agent that can access continuous, global market data, real-time financial news, and live sentiment data. In this way, we sought to provide users with relevant information that can help with live decisions.
Comprehensive architecture that includes retrieval and citations
There is a retrieval layer in the Xurve architecture that precedes answer generation.
We designed the system such that it can consider factors like recency, relevance, and comprehensiveness when retrieving data. The model’s job is not to invent evidence (as we will see) but to synthesize retrieved evidence based on these three factors.
Citations are another relevant part of the architecture. We designed our model such that the sources are not just cited, but that the system can associate the information provided with the underlying source (market news, market research, market data, and other financial information). In this way, users can logically move from the conclusion back to the sources.
Contextualized sentiment analysis
Instead of providing a simple bullish or bearish sentiment score for an asset, we took an approach that recognizes how different factors affect market sentiment.
For every news item, we identify all the assets affected and delineate whether that sentiment is bullish, neutral, or bearish. The same headline can be bullish for one company and bearish for the other. Also, the same macroeconomic news that is bullish in a low-interest-rate regime can be bearish in a high-interest-rate regime.
Thus, at any point in time, various factors can be bullish, bearish, or neutral for a given asset. Users can then interrogate all these factors through Xurve AI Analyst. Instead of just providing a general sentiment score, the conversational AI agent can help users understand the context behind each of the relevant factors driving market sentiment.
Lessons for AI builders
At IUX24, we made it our vision to provide institutional-grade market intelligence to retail investors. And at the heart of this vision is the Xurve AI Analyst.
For us, product differentiation was key to standing out in the bedlam of AI product rollouts. With everyone accessing the same foundation models, building a product with unique selling points that are relevant to users became an important driver.
And this is the lesson for all AI builders: the foundation model needs to become a smaller part of the product’s differentiation. Building an architecture that solves problems other models are ignoring or solves them in a unique or marginally better way is how you remain relevant.
For Xurve AI Analyst, this means integrating a market information layer, a market news and analysis layer, a retrieval layer, real-time market data, contextual sentiment analysis, and citation infrastructure.
Remember to also find your edge in the market and devote most of your product development time to it. That’s how you build a product that stands out.
