AI E-commerce Personalization Lifts Add-to-Cart Rate 34%
SpeedMVPs built an e-commerce personalization MVP for Threadline Commerce: a real-time AI personalization engine on their existing Next.js storefront, backed by pgvector on Supabase for product and user embeddings. A collaborative-filtering model trained on 18 months of order and clickstream data serves ranked recommendations per user in under 80ms through a FastAPI microservice, and an LLM-based intent classifier reranks their existing Algolia search. The engine lifted add-to-cart rate 34%.
SpeedMVPs built a real-time AI personalization engine on their existing Next.js storefront backed by pgvector on Supabase for product and user embeddings. A collaborative-filtering model trained on 18 months of order and clickstream data was deployed via a FastAPI microservice, surfacing ranked recommendations per user in under 80ms. Personalized search reranking was layered on top of their existing Algolia index using a lightweight LLM-based intent classifier that rerouted ambiguous queries to semantically similar in-stock products.
Book a free 30-minute callProject Overview
What We Built
- SpeedMVPs built a real-time AI personalization engine on their existing Next.js storefront backed by pgvector on Supabase for product and user embeddings. A collaborative-filtering model trained on 18 months of order and clickstream data was deployed via a FastAPI microservice, surfacing ranked recommendations per user in under 80ms. Personalized search reranking was layered on top of their existing Algolia index using a lightweight LLM-based intent classifier that rerouted ambiguous queries to semantically similar in-stock products.
The Challenge
- Threadline Commerce was running a single ranked bestseller list across their 80,000-SKU catalog: every visitor saw the same homepage grid regardless of browse history, cart, or purchase signals. Conversion on product detail pages sat at 1.8%, and their internal team had spent six months evaluating off-the-shelf recommendation SaaS tools that each cost $4,000–$12,000/month and required six-month onboarding contracts. They needed a production-ready personalization layer shipped before their autumn seasonal push.
Our Solution
- Real-time collaborative-filtering recommendations served via FastAPI with <80ms p95 latency
- pgvector product embeddings (text + image) for cold-start and cross-category discovery
- LLM-based search intent classifier reranking Algolia results for ambiguous queries
- Personalized homepage grid with A/B holdout group baked into the feature flag rollout
- Admin dashboard showing per-segment recommendation performance and override controls