AI Automation · Confidential client
AI Suite for a Multi-Brand E-commerce Operator
A coordinated AI surface across search, recommendations, conversational shopping, support, and lifecycle messaging — lifting conversion and deflecting most support contacts at a fraction of the prior unit cost.
- Conversion lift
- +14%
- Support deflection
- 62%
- Cost per support contact
- $3.40 → $0.15
- AOV lift (AI cross-sell)
- +9%
01 · Problem
What we were asked to solve
AI was being added to the storefront one vendor at a time — search by one tool, chat by another, recommendations by a third, lifecycle email by a fourth. Each surface optimized its own KPI, none shared customer signal, and unit economics worsened as the catalog and traffic grew. The brief was to replace the patchwork with a single substrate that learns from every interaction across the entire customer journey.
02 · Approach
How we built it
We built a shared AI substrate — one retrieval layer over the catalog, content, reviews, and customer history; one event stream; one evaluation harness — with thin per-surface adapters that replaced individual SaaS vendors in sequence. Search, recommendations, conversational shopping, support, and lifecycle messaging now read from the same signal, so wins compound rather than fragmenting across vendor dashboards.
03 · Architecture
The system, in five lines
- Unified retrieval over catalog, customer history, content, and reviews
- Hybrid semantic search — vector recall plus keyword, with LLM reranking
- Conversational shopping agent on PDPs and cart with full session context
- Personalized recommendations across home, PDP, cart, and lifecycle email
- Support automation with grounded answers, refunds and returns tools, and human handoff
04 · Stack
What it runs on
- Hybrid retrieval (vector + keyword + reranker)
- LangChain orchestration
- Conversational commerce agent
- Personalization engine
- Shopify · Klaviyo · Zendesk
- Real-time event streaming
Build something like this