OTS began by developing a framework to evaluate leading generative AI platforms including OpenAI, Gemini, and Grok to determine which would serve as the base model for our project. We systematically tested how each model performed when using techniques including context caching, advanced prompt engineering, and fine-tuning on manually evaluated training data to assess performance, cost, and accuracy. 

 

After developing several custom models and rigorous experimentation, a Gemini-based model emerged as most effective for our use case.  We then iterated further on this model, refining our approach through continuous testing to maximize extraction accuracy and consistency. This finally culminated in a data pipeline with the custom model at the helm.

​

After successfully leveraging our GenAI model to reliably extract the necessary data from the Word documents, OTS built several procedures to validate and format the data. The final step of the data pipeline annotated each furniture item’s data set with the categorization and special part information HFC desired. To ensure a seamless transfer into ServiceTitan, OTS implemented robust testing infrastructure to ensure all parts of the process worked as intended. This approach dramatically reduced manual effort, improved data integrity, and enabled a scalable solution for future growth.