AI Commerce Platform
AI-native search, merchandising, and personalisation layer for a multi-brand retail group operating across fourteen markets.
The situation
A multi-brand retail group operating across fourteen markets had a catalogue problem. Over 80,000 SKUs across six brands, managed in three separate systems, with search and discovery built on keyword matching from 2019. Conversion rates had flatlined despite significant paid spend. The brief was to build an AI-native commerce layer without replacing the existing platform infrastructure.
The constraint was hard: the group had committed to keeping its existing ERP and storefront architecture for thirty months. Any AI capability had to sit above the stack, ingest data from three systems in near-real time, and return results fast enough to not hurt Core Web Vitals. The diversity of the catalogue — fashion, home, and beauty — meant a single model approach would not hold.
A retrieval layer, not a replacement
Rather than rebuilding the commerce stack, we designed a retrieval and personalisation layer that intercepted search, browse, and recommendation requests, enriched them with AI-generated signals, and returned ranked results back to the storefront. Each brand got a domain-specific model tuned on its own purchase and browse data.
- Per-brand embedding models trained on eighteen months of purchase data
- Real-time retrieval pipeline processing 4,000 requests per minute at peak
- Merchandising rules engine allowing manual overrides with AI ranking preserved
- A/B testing framework embedded from day one to isolate conversion lift
Eight weeks from architecture to production
Data Audit & Architecture
Mapped all three catalogue systems, defined the ingestion pipeline, and designed the embedding architecture before writing a line of production code.
Model Training & Integration
Trained per-brand models on historical data, built the retrieval API, and integrated with the existing storefront via a thin middleware layer.
Instrument & Optimise
Deployed with full observability — latency monitoring, ranking evals, and conversion tracking — and ran four weeks of A/B tests to validate lift before full rollout.
+38% conversion. 11× catalogue throughput.
Conversion on the primary brand lifted 38% within six weeks of full rollout. Catalogue throughput — the number of SKUs surfaced to relevant customers — increased 11× against the keyword baseline. Latency added by the AI layer averaged 18ms at the 95th percentile, below the agreed threshold.
More programmes
Brand Identity System
Strategic repositioning, identity system, and editorial platform for a multi-market financial group ahead of a $400M raise.
Venture Build
Co-founded and built a fintech operating system for SMBs from memo to commercial product, including identity, app, and growth.
Growth System
A measurement-first growth system across paid, lifecycle, content, and SEO for a $90M ARR enterprise SaaS approaching IPO readiness.
Ready to build
something real?
One conversation. One operating contract. One accountable partner from brief to launch and beyond. Tell us what you're building.