NeuralCart

Challenge

A retail platform processing high daily transaction volumes was relying on a rules-based recommendation engine built five years earlier. Suggestions were static category-level outputs that ignored individual browsing behaviour, purchase history, and real-time stock availability. The gap between what customers wanted and what they were shown was costing the business meaningful revenue per session.

Approach

We audited the platform's existing event data and found that while browsing and purchase signals were being logged, none fed back into the product experience. We designed a two-phase approach: fine-tuning a language model on the product catalogue and customer interaction history to build semantic understanding of product relationships, then building a lightweight inference layer to serve personalised recommendations without adding latency to the checkout flow.

Solution

The system combined a fine-tuned GPT-4 embedding model with a retrieval-augmented ranking layer that scored candidates against real-time inventory and margin data. A React component rendered recommendations at three touchpoints — product detail pages, cart review, and post-purchase. The inference pipeline ran on AWS Lambda with sub-150ms response times under peak load. Accepted and rejected recommendation signals continuously refined future outputs.

Results

• +47% uplift in average order value measured over 30 days
• 87% of served recommendations clicked or added to cart
• Deployed end-to-end in 6 weeks from kickoff to production
• Zero performance degradation observed during peak traffic events