RecoEngine Personalization

RecoEngine is a production personalization platform built for a European fashion and lifestyle retailer operating across 14 markets. The brand needed a recommendation layer that felt native to the storefront — not a generic widget — while handling millions of SKUs, seasonal inventory swings, and strict latency budgets on mobile.

I led the ML workstream from discovery through SageMaker deployment, partnering with product, data engineering, and merchandising stakeholders to align model behavior with commercial goals.

The client's legacy rule-based recommendations were static and underperformed on long-tail catalog items. Category managers spent hours manually curating collections. Click-through rate on recommendation slots stagnated at 1.8%, and the data science team lacked a reproducible experimentation pipeline.

Key constraints included sub-100ms inference at p95, GDPR-compliant user profiling, and the need to explain ranking decisions to merchandising teams.

We implemented a two-tower retrieval model for candidate generation paired with a gradient-boosted ranking layer fed by real-time behavioral features stored in Redis. Offline training ran on PyTorch with daily feature store refreshes; online serving used AWS SageMaker endpoints behind a lightweight API gateway.

An A/B testing framework allowed merchandising to compare model versions against holdout groups with automated guardrails on revenue and latency regressions.

- 22% lift in recommendation slot CTR within 8 weeks of launch - 14% increase in revenue per session on personalized home feeds - p95 inference latency held at 78ms under peak traffic - Merchandising time on manual curation reduced by ~40 hours/month

Owned model architecture, feature engineering, SageMaker deployment, and experimentation framework. Mentored one junior data scientist and documented MLOps runbooks for the internal platform team.