Global Sportswear Brand

Product search on large e-commerce catalogues has a known failure mode: filters and keywords work for people who already know what they want. They fail for shoppers who want guidance — "what shoe would work for trail running on wet terrain" doesn't map cleanly to 14 filter dimensions.

Proving that conversational AI could handle that mode of discovery at catalogue scale required solving three things:

• Intent interpretation across ambiguous language. "Something for the gym but not too bulky" has to be matched against product attributes that don't use that language.
• Multi-turn conversation state. A real discovery conversation is refinement over multiple exchanges, not a single query.
• Speed. A shopping experience that feels like an interrogation doesn't convert.

We built a conversational product discovery prototype in 4–6 weeks:

• Pinecone — vector embeddings of the full product catalogue, enabling semantic matching against natural language queries
• LangGraph — multi-turn conversation state management, tracking context across the session so follow-up questions refine results rather than restart the search
• Next.js — frontend interface designed to feel like messaging, not search

We validated the prototype with user testing, measuring task completion, satisfaction, and conversion intent against the existing filter-based experience.

The prototype demonstrated that conversational discovery outperformed keyword search for open-ended intent — shoppers who didn't know exactly what they were looking for found relevant products faster.

The brand advanced it to their production roadmap. That's the validation that matters.