MERN Storefront Rebuild + AI Merchandising

The Problem

A DTC brand at 50K MAU was running a 6-year-old PHP storefront that had been patched so many times it was held together with duct tape. TTFB of 2.1 seconds, a checkout flow that dropped 38% of users at step 3, and a search system that returned irrelevant results for anything beyond exact product names.

The founders didn’t just want a port — they wanted AI-powered product discovery. But they also couldn’t afford a 6-month rewrite.

I delivered both in 12 weeks.

Architecture

Frontend — Next.js 14 (App Router)

Server Components for product pages (SEO + speed)
Client Components only for cart, search, and interactive UI
Incremental Static Regeneration for collection pages (30s revalidation)
Vercel Edge Network for global CDN

Backend — Node.js + Express API

Clean REST API separating concerns from the frontend
MongoDB with proper indexing (the old PHP app had none)
Redis for session data and cart persistence
Stripe for payments (migrated from legacy Braintree)

AI Merchandising — Embeddings-based Discovery

The search and product recommendation system uses OpenAI text-embedding-3-small:

All 4,200 products embedded at ingest (title + description + tags)
User search queries embedded at runtime
Cosine similarity search against product embeddings
Results ranked by semantic relevance × inventory × margin score

This replaced the old exact-match keyword search with semantic understanding. “Casual weekend wear” now returns relevant products even if those exact words don’t appear in any product description.

The Migration Strategy

I ran the new and old systems in parallel for 6 weeks with a feature flag controlling which experience each user sees. This let us:

A/B test conversion rates before full cutover
Migrate order history and customer data incrementally
Roll back instantly if anything broke

Results

TTFB: 2.1s → 180ms (92% reduction)
Conversion rate: 3.4× lift on semantic search vs. keyword search (A/B tested)
Cart abandonment at step 3: 38% → 14%
Search-to-purchase rate: 2.1% → 7.8%
Timeline: 12 weeks from kickoff to production cutover

What I’d Do Differently

I underestimated the data migration complexity. Product variant data in the old DB was a nested string blob — parsing it took a full week I hadn’t scoped. Always audit the source data schema before estimating a migration.