Fetch, the rewards app that lets users scan receipts for points, now processes over 80 million receipts per week — that’s hundreds of scans per second at peak traffic. The company needed to extract text, parse purchase data, and structure it fast enough to reward users instantaneously. In 2021, Fetch set out to overhaul its ML pipeline, and by late 2022 the company had built, trained, and deployed more than five machine learning models using Amazon SageMaker, cutting processing latency by 50 percent.
The company leans heavily on SageMaker Model Training and SageMaker Processing, running custom models on multi-GPU instances. According to Sam Corzine, ML engineer at Fetch, “Amazon SageMaker is a game changer for Fetch. We use almost every feature extensively.” The team also credits Hugging Face’s AWS Deep Learning Containers for simplifying model deployment. “Using the flexibility of the Hugging Face AWS Deep Learning Container, we could improve the quality of our models,” Corzine said.
Deployment friction is nearly nonexistent now, which has had an unexpected cultural payoff. Applied scientist Alec Stashevsky noted that “there’s little friction for us to deploy models. Basically, we don’t have to think about it.” A new intern reportedly deployed a model by his third day on the job — a stat that says more about the tooling than the intern. Fetch also uses SageMaker Inference Recommender to automate load testing and model tuning.
The accuracy gains are substantial: Fetch improved its document-understanding model by 200 percent. That’s a meaningful leap for a company whose entire business model hinges on correctly parsing crumpled, faded, inconsistently formatted receipts. Pairing Hugging Face’s model ecosystem with SageMaker’s managed infrastructure gave Fetch a blueprint that other AWS-native companies processing high-volume unstructured data — think logistics firms or expense management startups — could reasonably copy. The question now is whether Fetch can sustain those accuracy gains as receipt volume keeps climbing.
