Roundforest Case Study
One Goal.
They wanted to use Generative AI to improve 3 key areas:Goal 1: IdeationExisting translators lacked contextual and semantic understanding. Their literal translations didn’t always return the most popular search terms.Goal 2: List generationWhen multiple search filters were translated and applied, the current LLM struggled to apply filters correctly when asked by consumers.Goal 3: Related listsRoundforest wanted to generate lists of related products, based on users search terms.
Improve price sensitivity to maximize in-game store monetization.
By deploying Generative AI, we opened the door to add more context and depth to prompts, increasing the volume of responses and applying multilingual prompting techniques using the Sander Schulhoff approach.
We used GPT-4 to return structured Json outputs and combined a prompting approach of Chain of Thought and Few Shot prompting to apply filters correctly and break categories down meaningfully.
We delivered a Generative AI solution for an AI-powered personal shopping tool. Its existing LLM model was lacking contextual understanding into search terms, especially when translated literally into other target market languages. This limited the potential of filtered searches, which many consumers rely on. Upselling and cross-selling was also limited, because of the inability to generate relevant ‘related products’ based on search results.
Results & outcomes:- 35% → 76% translation accuracy- Products now offered in multiple countries- 3X list relevancy- Increased personalization at scale
