unthinkable
Case study · ETP, Singapore · retail software provider

ETP's retail customers get anomaly alerts and product recommendations as soon as they subscribe, with no manual setup.

ETP, which serves 300+ brands and 35,000+ stores, wanted each retailer that subscribes to have its own trained models working quickly and reliably.

SingaporeRetail analytics & trade promotions

What stood in the way

What stood in the way.

Each tenant's ML models were set up and deployed by hand, which was slow, error-prone and inefficient.

Manual setup could not keep pace with a growing customer base and made onboarding new tenants slow.

What we built

What we built.

  • The client's anomaly detection notebook turned into code using an extended isolation forest model, with MLflow for experiment tracking and as the model registry.
  • Weekly training for every tenant scheduled on Apache Airflow, with batch predictions every 20 minutes stored in GCP buckets and old data cleared within an hour.
  • Training and deployment of three recommendation types: items for a user, users for an item, and item to item.
  • Retraining frequency chosen by each company, with a daily job picking which models to train.
  • A separate backend engine for each new tenant in its own namespace, deployed automatically through Airflow and Jenkins, serving recommendations through APIs.
  • Apache Airflow
  • MLflow
  • Jenkins
  • Google Cloud Platform
  • Extended isolation forest
What it shows

What this work shows about us.

AI that runs in production

The models already existed; the work was making them train, deploy and serve every tenant on their own, which is where most ML projects stall.

Each retailer kept apart

A backend per tenant in its own namespace keeps every retailer's data isolated.

Want to automate machine learning training and deployment like this? Tell us what you want to change. Our AI assistant, Ask Unthinkable, helps you define your requirements, then connects you with the right team.

Start your plan ↗
If we built it today

If we built it today, with AI.

What we would add now, moment by moment.

Finding it

An agent would use the recommendation APIs to suggest items to a shopper in conversation, explaining why each one fits.

Choosing in store

An agent would explain an anomaly alert to a store manager in plain words and say what to check first. The manager decides what to do.

Ask Unthinkable

Want this for your business?

Tell us what you run or what you're building. Our AI assistant, Ask Unthinkable, helps you define your requirements, then connects you with a senior AI engineer.

Send us a note instead

Required

Thanks. A senior AI engineer will reply to you.
u
Ask UnthinkableClarity first, then the right team
Your conversation

An AI agent. This prototype uses prepared answers and sends nothing.