Context:
To increase relevance and reach of Amazon Prime value, as well as awareness of how Prime benefits can apply to customer’s interests and orders, the Acquisition team wants to design a curated recommendation widget powered by generative AI to surface products and digital content that align with customers’ recent browsing and purchasing activity. In order to do so, the AI model will need guardrails around content creation—specifically voice, tone, and style—to ensure auto-generated messaging is tailored to the customer while remaining on-brand.
• I developed a set of contextual content design and messaging rules to train the AI model to generate headlines that capture the Prime member’s attention without “selling” them anything
• I created a heuristic to train the AI model on what aspects of member’s profile and activity to combine into messaging that inspires and intrigues while clearly connecting value back to Prime
• I trained the AI model through reinforcement learning and iterated on guidelines whenever certain scenarios (e.g. a customer who is shopping on behalf of someone other than themselves) produced hallucinations
Customer personas (left) & personalized AI-powered inspiration (right):
🔙