Ecommerce ops · Production

Machine Learning at Shopify: product classification, fraud detection, Sidekick merchant assistant, and GMV forecasting

The problem

Shopify must optimize commerce for millions of merchants simultaneously across product metadata quality, fraud prevention, merchant guidance, capital access, and search relevance — a continuously expanding optimization problem at massive scale.

Workflow diagram · grounded in source
1
Product classification and enrichment
ai_action
“We use finetuned Qwen multimodal models for classifying and enriching the metadata of each product uploaded into the system. This entails hundreds of millions of inferences a day.”
2
Transaction fraud assessment
ai_action
“We assess fraud of every transaction using very fast risk models.”
3
Sidekick merchant assistant
ai_action
“We use a combination of fine-tuned LLaMa models and large general models, combined with refined MCPs, to build Sidekick, our loyal multi-purpose merchant assistant.”
4
Product embeddings for discovery
ai_action
“We rely on finetuned Nomic embeddings for vector representations of the many billions of products in our catalog. These embeddings empower product discovery systems including search and recommendations.”
5
Merchant GMV forecasting
ai_action
“At the heart of our risk assessment tools, we developed a tabular transformer based model for forecasting merchant GMV.”
6
Customer and merchant behavior modeling
ai_action
“Sequence based foundational models are key for understanding merchant and customer actions at detail and determining what would be the best move given their current objectives. For this, we are actively experimenting with the HSTU archit…”
7
Search query rewriting
ai_action
“Query rewriting increases the chance of offering the customer what they are really looking for. This requires very small and very smart language models producing better queries at request time.”
Reported outcome

Shopify has multiple ML systems in production including product classification running hundreds of millions of inferences per day, fraud assessment on every transaction, a multi-purpose merchant assistant (Sidekick), vector embeddings powering search and recommendations across billions of products, and GMV forecasting for merchant capital access.

Reported metrics
Product classification inferences per dayhundreds of millions of inferences a day
Reported stack
LLaMa modelsGCPNebiusCentMLToloka
Source
https://shopify.engineering/machine-learning-at-shopify
Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

Shopify has multiple ML systems in production including product classification running hundreds of millions of inferences per day, fraud assessment on every transaction, a multi-purpose merchant assistant (Sidekick),…

What tools did this team use?

LLaMa models, GCP, Nebius, CentML, Toloka.

What results were reported?

Product classification inferences per day: hundreds of millions of inferences a day (source-reported, not independently verified).

How is this ecommerce ops AI workflow structured?

Product classification and enrichment → Transaction fraud assessment → Sidekick merchant assistant → Product embeddings for discovery → Merchant GMV forecasting → Customer and merchant behavior modeling → Search query rewriting.

Related ecommerce ops cases