How We Built Nike's Personalized AI Shopping Assistant in Just 3 Weeks
In just three weeks, we delivered for Nike a generative AI-powered shopping assistant that did more than recommend products–it created a dynamic, personalized shopping experience. The assistant engaged users in conversation, adapted recommendations to their preferences, shopping history, and profile in real-time, and improved with use.
Introduction
With thousands of product SKUs and data from over 170 million loyalty program members, Nike envisioned a hyper-personalized, conversational shopping experience.
One that felt like a natural, one-on-one conversation and made customers feel understood and supported, by surfacing relevant products in real-time and at scale, in an intelligent, dynamic, and non-intrusive way.
Our Approach
The proof-of-concept had to be built and deployed quickly, to demonstrate value and win internal stakeholder support.
Build & Test
We built the initial proof-of-concept in just eight days. We used publicly available product data and synthetic user data for the core build, and anonymized data for testing. This allowed our team to begin building on day one, and start testing and iterating with client feedback by week 2.
Deploy
Vercel and Next.js are our preferred choice for fast and secure, zero-configuration deployments. Vercel’s global edge network makes it easy to scale dynamically and deliver fast, high-performance applications.
Control
Airbender is an open-source platform we created to enable real-time monitoring of AI products–complete with audit logs and a kill switch. It empowers teams to monitor, control, and manage the AI’s operations without developer support.
We don't just talk about doing things, we actually do it.
In three weeks, we built and deployed a conversational shopping assistant.
Watch Ethan, a Nike member who’s looking for golf attire, and Sophia, who’s looking for a yoga outfit, interact with the AI shopping assistant.
Our mixed embeddings model leveraged three data sources: real-time conversation, member data, and enriched product catalogue embeddings.
We fine-tuned the model to return relevant products weighted by the user’s known preferences. This made it possible to recommend highly personalized products, such as high-waisted, postpartum yoga pants in the customer’s preferred leggings colour for someone who’s recently purchased products in Nike’s maternity collection, for example.
Enriching product data using a large language model made recommendations more accurate and relevant.
Read more about using AI and multimodal language models to enrich e-commerce product data and enable smarter experiences.
Leveraging an LLM for summarization reduced token use, cutting costs and allowing the AI assistant to handle more interactions within the available context window.
We implemented a retrieval-augmented generation (RAG) with OpenAI’s latest text generation model to ensure the AI assistant accurately recommended available Nike products without hallucinating.
Airbender provides a dashboard for non-technical admins to monitor, maintain, and manage the AI assistant in both production and testing. They could track real-time usage, review customer interactions and AI assistant responses via detailed audit logs, collect user feedback, create automated workflows with predefined triggers (e.g. terminate interactions with malicious users, escalate support requests), or disable the AI assistant altogether via a kill switch–all without requiring developer involvement.
These steps help ensure the AI assistant is responsive, safe, and aligned with evolving needs, even after our engagement.
Highlights
- In three weeks, we delivered Nike’s AI shopping assistant, unlocking the potential of their vast membership data and extensive product catalog
- Within eight days, we built the initial PoC by leveraging public and synthetic data for the core build
- Airbender empowered non-technical staff to monitor, audit, and manage the AI assistant without developer resources
Conclusion
For e-commerce brands looking to securely deploy AI-driven products, integrating a stack like Vercel, Next.js, and Airbender with generative AI capabilities can enable you to move at the speed of the market, without sacrificing control, privacy, or performance.