The Challenge: 

Tier-1 global RoRo logistics operations generate massive volumes of work orders logged across dozens of terminals — sparse, short-text entries, often just two to four words, written in English, Spanish, and French, saturated with localized jargon and cryptic abbreviations. The mission: classify every work order into a unified corporate taxonomy of 400 classes taxonomy(240+ Active) distinct active classes, in real time, at enterprise scale.

Standard RAG-LLM architectures were evaluated and failed — stalling at 70% accuracy. Three compounding factors made this problem exceptionally hard: 

1. generative models lack the domain grounding to resolve cryptic terminal abbreviations; 

2. hard taxonomy boundaries require precise discrimination, not generation — LLMs hallucinate at these edges; 

3. severe class imbalance across 240+ active classes meant the model saw thousands of examples of common work orders and almost none of rare but operationally critical ones. 

No off-the-shelf architecture handles all three simultaneously. The client's own expert-estimated ceiling for human performance on this dataset: 80%.

The Solution: 

TQW Solutions applied a methodology no generalist AI vendor could replicate: transferring deep expertise in high-frequency financial time-series classification — where signal is sparse, noise is extreme, and precision is non-negotiable — into the maritime domain.

The result was a custom semantic classifier with calibrated confidence: a specialized ensemble model with smart sampling to manage class imbalance, engineered specifically to succeed where generalized RAG-LLMs fail. Fully deployed in a production Azure environment.

The Mission Impact: 

Production accuracy: 96.3% at 80% coverage and 87.8% at full coverage — surpassing the client's expert-estimated performance ceiling of 80% and every LLM baseline evaluated.

By integrating the model as classifier directly into the corporation's central enterprise system, TQW eliminated operational data silos and delivered a real-time Common Operating Picture (COP) across operations, finance, and executive leadership — transforming chaotic field data into reliable, audit-ready business intelligence.

By integrating the model as a recommendation engine, operators are guided toward selecting the correct service codes during work order creation. Adapting disparate systems into a unified enterprise model directly at the point of data entry enables a gradual, non-intrusive transition toward normalized coding practices.