LLM Results Analyzer: Unified Analysis Framework

Executive Summary

The LLM Results Analyzer integrates outputs from the Custom Clarification Project (theoretical HS code data) and Custom Optimization Project (real-world product descriptions) into unified classification recommendations. This solution tackles the challenge of leveraging both theoretical compliance knowledge and practical trade data to provide accurate and well-justified HS code classifications. The system features a unified endpoint that simultaneously sends requests to both LLM systems, with a third specialized LLM analyzing and synthesizing the results from both sources, providing users with recommendations that consider both legal compliance requirements and practical trade precedent.

Introduction

The Custom Clarification Project and Custom Optimization Project created two powerful but separate systems for HS code classification. The Custom Clarification Project provided access to theoretical HS code definitions, official regulations, and legal frameworks essential for compliance accuracy. The Custom Optimization Project offered insights from 65.5 million real-world product descriptions, reflecting actual trade practices and classification precedents. While both systems delivered value independently, users faced the challenge of determining how to best utilize both sources of information for optimal classification decisions. The LLM Results Analyzer addresses this integration challenge through an orchestration framework that automatically combines insights from both systems and provides unified, well-reasoned recommendations.

The Problem

The existence of two separate classification systems created an integration challenge that limited the potential value both systems could deliver when used together. Users found themselves with access to theoretical knowledge through the Custom Clarification Project and extensive practical data through the Custom Optimization Project, but no efficient way to combine these resources. The primary challenge was workflow complexity - users had to manually submit the same query to both systems, wait for separate responses, and then attempt to reconcile potentially different recommendations. Inconsistent results between the two systems created confusion and decision paralysis, with users lacking guidance on how to resolve these discrepancies or determine which recommendation to prioritize. The manual comparison process was prone to human error and inconsistency, and the time required to use both systems effectively became a barrier to adoption.

The Solution

The LLM Results Analyzer was developed as an orchestration framework that automatically combines the strengths of both the Custom Clarification and Custom Optimization projects into a single, unified user experience. When users submit a classification request to the unified endpoint, the system immediately dispatches identical queries to both the Custom Clarification and Custom Optimization endpoints simultaneously. This parallel processing approach ensures optimal efficiency by eliminating sequential wait times while gathering information from both theoretical and practical perspectives. Once both systems return their results, a specialized analyzer LLM processes the combined outputs using prompt engineering designed specifically for HS code classification synthesis. This analyzer LLM understands the nuances of both regulatory compliance and practical trade classification, enabling it to identify areas of agreement, explain discrepancies, and provide reasoned recommendations that consider both perspectives. The solution maintains full transparency by preserving the original outputs from both underlying systems while adding the synthesized analysis layer.

Results

The LLM Results Analyzer delivered improvements in both user experience and classification accuracy by successfully integrating the strengths of both theoretical and practical HS code classification approaches. The unified endpoint eliminated the complexity and time overhead associated with manually using two separate systems. The parallel processing architecture ensures that total response time is only marginally longer than the slower of the two underlying systems. The analyzer LLM successfully identifies and explains discrepancies between theoretical and practical approaches, providing users with clear guidance on how to resolve classification conflicts. Classification confidence improved as users now have access to justifications that consider both legal compliance requirements and real-world trade precedent. Adoption rates increased substantially as the simplified workflow eliminated the time and complexity barriers that previously prevented users from fully leveraging both systems.

Conclusion

The LLM Results Analyzer successfully solves the integration challenge that limited the combined potential of the Custom Clarification and Custom Optimization projects. By creating an orchestration framework that automatically synthesizes theoretical and practical HS code classification approaches, the solution delivers comprehensive analysis while maintaining user-friendly simplicity. The parallel processing architecture and specialized analyzer LLM demonstrate how AI can be used not just to automate individual tasks, but to combine multiple knowledge sources for better decision-making support. This project demonstrates the potential of system integration in complex classification environments. By solving the challenge of combining theoretical knowledge with practical experience, the LLM Results Analyzer enables organizations to make more informed, confident, and accurate HS code classification decisions that satisfy both compliance requirements and operational efficiency needs.

Figure 1.1. Graphical interface of the customs HS code classification system showing the outputs of the Custom Clarification API (left) and the Custom Optimization API (right). The left panel displays the system's automated analysis of a product description, based on the theoretical data with recommended HS codes. The right panel lists potential optimized product descriptions from the 65.5 million dataset. This figure illustrates how the system combines both results and analyzes final HS code suggestions. For more, you can watch the video at the link.