The challenge
Underwriting processes require collecting and analyzing various client-related data points, such as:
- Ownership structure of companies
- Major corporate events
- Revenue trends
- Stock value insights
Traditionally, this required extensive manual web-surfing and cross-referencing information across multiple online sources. The challenges included:
- Repetitive Effort: The same types of queries were conducted repeatedly for different clients.
- Interconnected Data: One piece of information often led to others, requiring a multi-step investigation.
- Time-Intensive: On average, each case took 20 minutes of manual work to gather and cross-check the required data.
- Scattered Information: Synthesizing insights from multiple sources into a cohesive summary was both challenging and prone to errors.
- LLM Limitations: Large language models are inherently limited to knowledge available during their training time. This knowledge gap made it crucial to dynamically incorporate up-to-date resources.
The client chose VirtusLab to explore whether integrating LLMs into this process could streamline information gathering and summarization, ultimately increasing efficiency and accuracy.
The solution
VirtusLab worked closely with the client to develop a POC that showcased the use of LLMs in the underwriting process. The solution addressed key challenges by implementing the concept of LLM Agents, enabling the automation of information gathering and summarization. The POC was developed using the LangChain framework and involved three key components:
- Agent: Served as the decision-making engine, managing the steps needed to fulfill a query based on LLM-generated instructions.
- Tools: Modular components designed to execute specific tasks, such as performing Google searches or retrieving online data.
- AgentExecutor: Orchestrated the process, calling the agent, gathering input from tools, and delivering consolidated results.
