Real-Time Personalisation Engine for 5M Monthly Users
Generic Experiences Costing Crores in Lost Revenue
The brand's e-commerce platform served the same product catalogue to every visitor regardless of browsing history, purchase behaviour, or real-time intent signals. A/B tests showed personalised experiences could lift conversion by 20%+, but the legacy recommendation engine — a batch-updated collaborative filter — refreshed only once every 24 hours and couldn't respond to in-session behaviour.
Event-Driven ML Personalisation on Google Cloud
SkySync replaced the batch recommendation engine with a real-time event-driven architecture on Google Cloud. Every user interaction — clicks, dwell time, add-to-cart, search queries — is streamed into Pub/Sub, processed by a custom two-tower neural network, and served as personalised recommendations in under 80ms.
Real-Time Event Pipeline
Instrumented the frontend with a lightweight event SDK streaming to Google Pub/Sub. Dataflow jobs process 2M+ events per hour with sub-second latency.
Two-Tower Recommendation Model
Trained a two-tower neural network on 18 months of purchase and browse history. The model updates embeddings every 15 minutes via Vertex AI Pipelines.
Low-Latency Serving Layer
Deployed recommendations via Cloud Run with Redis Memorystore caching, achieving p99 latency of 78ms — down from 1.4 seconds on the legacy engine.
Performance & CDN Overhaul
Migrated static assets to Cloud CDN with edge caching, implemented lazy loading and image optimisation, cutting mobile load time from 4.2s to 1.4s.
Real-Time ML on Google Cloud
Google Pub/Sub + Dataflow
Real-time ingestion and processing of 2M+ user events per hour
Vertex AI Pipelines
Automated two-tower model retraining every 15 minutes
Cloud Run + Redis Memorystore
Sub-80ms recommendation serving with in-memory caching
BigQuery + Cloud Bigtable
User feature store and historical event warehouse
Google Cloud CDN
Global edge caching for static assets and API responses
Cloud Monitoring + Looker
Real-time A/B test dashboards and model performance tracking
Revenue Impact in 90 Days
"We'd been talking about real-time personalisation for two years. SkySync shipped it in five months and the revenue impact was visible within the first 90 days. The architecture they designed will scale us to 50 million users."
Ready to achieve similar results?
Talk to our architects about your cloud and AI challenges. We'll scope a proof-of-value in two weeks.