Meridian Commerce operates a multi-brand fashion and homewares platform with 2.4M monthly visitors. Despite 18 months of A/B testing, conversion rate had plateaued at 3.9%. The core problem: A/B tests optimise for the average visitor. Every winning treatment improved one segment while hurting another, making universal improvement impossible without per-visitor intelligence.
CVR Plateau at 3.9%
One-Size Personalisation
68% Cart Abandonment
240M Daily Events Wasted
240M Events/Day
180 Features/Session
12ms Inference
Rules-Free Decisions
Per-Session Products
Personalised Pages
Redis
OpenAI
MLflow
Kubernetes
Apache Kafka
Apache Flink
Flink pipeline processing 240M+ daily events. Computes 180 behavioural features per session in real time: scroll depth, dwell time, category affinity, and price sensitivity signals.
Gradient-boosted model trained on 18 months of session data. Predicts likelihood to purchase (0–1), price ceiling estimate, and preferred category cluster — updated on every page event.
Rules-free offer system: when intent crosses a threshold, the engine selects the optimal intervention based on predicted price sensitivity, inventory margin, and offer history.
Category pages re-rank in real time based on the visitor's category affinity vector and collaborative filtering from the closest behavioural cohort — zero cold-start latency.
Every decision and outcome feeds back into model retraining via a daily Feast feature store refresh. The model improves weekly — A/B holdout groups continuously measure ongoing lift.
Conversion Lift
Added Revenue
AOV Increase
Decision Latency
