Commerce as an Optimization Problem
Shopify frames its platform as a giant optimization challenge. Each merchant runs a distinct business with its own definition of success, but the underlying infrastructure—product creation, discoverability, transactions, capital access—must serve them all. Machine learning is the tool Shopify uses to break that infinite problem into finite, tractable pieces. Some optimizations are deeply personalized to individual merchant goals; others improve the platform for everyone simultaneously.
Two decades of merchant data underpin this work. Shopify's historical dataset captures the trajectories of millions of businesses at high resolution, encoding practical knowledge about how companies operate and the decisions they make. That data is the raw material for the company's ML systems.
Where ML Is Already at Work
Shopify's ML investments span several core areas of the commerce lifecycle:
- Product understanding: Fine-tuned Qwen multimodal models classify and enrich metadata for every product uploaded to the system, a workload that runs hundreds of millions of inferences per day. The models determine what a merchant sells and which product characteristics deserve highlighting.
- Fraud detection: Risk models score every transaction in real time. Shopify cites Feature Space's work as inspiration for these fast inference systems.
- Merchant assistance: Sidekick, Shopify's assistant, combines fine-tuned LLaMa models with larger general-purpose models and refined MCPs (Model Context Protocols) to help merchants use the full range of platform capabilities.
- Product representations: Fine-tuned Nomic embeddings generate vector representations for billions of catalog products, powering search and recommendation systems.
- Capital and risk: To offer merchants financing on good terms, Shopify forecasts long-term GMV (gross merchandise value). Its risk assessment tools rely on a tabular transformer model drawing on ideas from tabular transformers, FT-Transformer, and relevant prior work.
- Behavior modeling: Shopify is experimenting with the HSTU architecture to build sequence-based foundational models that capture merchant and customer actions in detail, then predict the best next move given current objectives.
- Query rewriting: The search team is exploring lightweight, fast language models that rewrite customer queries at request time to improve match quality, guided by work such as this paper.
These use cases represent a sample, not the full inventory, of Shopify's ML work.
Infrastructure and Tooling
Shopify's ML infrastructure strategy centers on flexible access to compute. The company partners with GCP as its primary infrastructure provider and works with neo-cloud providers such as Nebius for large training clusters. Internal tooling abstracts away the underlying cloud vendor, so teams can move between providers to access the latest GPUs as they become available.
For workloads that demand high throughput or extremely low latency, Shopify works with partners like CentML, which bring specialized expertise in accelerating GPU computations. Data quality gets equal attention: Shopify supplements its internal data operations with services from Toloka to ensure high-quality training data is available quickly.
Shopify also maintains early access to frontier models from major labs. These models serve multiple purposes: annotations, evaluations, and direct deployments for specific use cases such as Sidekick. They're also attached to coding assistants to accelerate prototyping and development. On top of external tools, Shopify is building an internal collection of ML tooling to streamline the practice of machine learning.
What Shopify Looks for in ML Hires
The company frames its ideal candidates as builders, tinkerers, and explorers—people who are curious and absorbed in their craft. Credentials matter less than demonstrated capability: Shopify wants researchers comfortable in production systems and engineers who can navigate the arXiv. The company explicitly discourages specialization into narrowly defined roles.
