ROLFtech is normalizing a service directory containing 474,712 entries accumulated over 35 years. The dataset includes duplicates, typos, and inconsistent terminology, which complicates vehicle maintenance scheduling and data analytics for the company's Dealer Digital Platform.
The team initially attempted to use full-text search, but the approach proved costly and inefficient, requiring an estimated six months of work for four backend developers. The project shifted to using LLMs, including Claude Opus, Sonnet, and GPT models, to map service operations to a standardized tree of vehicle components.
Following deduplication, the number of unique service operations was reduced to 247,595. A team of 21 internal experts, including master consultants and warranty engineers, validated the LLM-generated mappings. In the first pass, experts confirmed 40.1% of the automated assignments.
To improve accuracy, the team implemented iterative feedback loops, allowing experts to correct misclassifications and suggest proper nodes. By providing relevant node suggestions during the review process, the team maintained a processing speed of approximately 2,000 operations per expert per day.
As of the current project stage, 96.13% of the deduplicated service operations have been successfully mapped to verified vehicle components. The project aims to integrate this cleaned data into the Flora platform to streamline service center operations and client interactions.
