RetailTech AI Personalisation Engine — UK Fashion Retailer

Personalisation ML Architecture

Collaborative filtering model (Matrix Factorisation using implicit feedback — browse, add-to-bag, purchase, return) deployed on AWS SageMaker.

Feature engineering

• user recency/frequency/monetary (RFM), category affinity scores, brand affinity, price sensitivity, seasonal behaviour patterns.
• Real-time inference: SageMaker real-time endpoint (p95 &lt
• 80ms) serving Next.js product recommendation components.
• UK GDPR Consent-

ICO cookie guidance

• personalisation based on browsing history requires PECR consent (not legitimate interest).
• Consent-tier architecture: Tier 1 (no consent) → session-based recommendations (no persistent profile), Tier 2 (analytics consent) → cross-session recommendations using anonymised ID, Tier 3 (personalisation consent) → full ML personalisation with named account profile.

Consent withdrawal

profile deletion within 24 hours.

Responsible AI Framework

Non-discrimination audit: personalisation model tested for proxy discrimination (price sensitivity score correlated with postcode-inferred deprivation index — removed as feature).

Transparency

• "Why am I seeing this?" explanation on every recommendation (driven by: your recent views, similar customers, trending in your size).
• Right to object to automated decision-making: opt-out of ML personalisation returns to editorial curation.

A/B Testing and Revenue Attribution

Multi-armed bandit (not pure A/B) for faster convergence: Thompson Sampling allocation between model variants.

Revenue attribution

ML recommendations tagged with recommendation_id — tracked through add-to-bag and purchase events via server-side analytics (not client-side — PECR compliant without additional consent).

Statistical significance

Bayesian approach (probability of being best, not p-value).