Fetch, the consumer rewards app with 18 million monthly active users, was flying blind. Its third-party receipt-processing AI worked—sort of—but gave the company zero visibility into how it extracted data from the 11 million receipts flowing in daily. That mattered because Fetch’s brand partners wanted granular insight into promotion engagement, and the company couldn’t deliver. “That solution was a black box and we had no control or insight into what it did,” said Boris Kogan, computer vision scientist at Fetch. “We just got results we had to accept.”
So Kogan did what any engineer would do when handed a black box: he built a replacement. Fetch assembled an in-house ML team and committed to a 12-month timeline on AWS. They finished in eight months. That speed came partly from AWS infrastructure, but the real accelerant was Hugging Face. Through the Hugging Face Expert Acceleration Program on the AWS Marketplace, Fetch got hands-on guidance—not project execution, but knowledge transfer. Yifeng Yin, a Hugging Face machine learning engineer, put it plainly: “They wanted to learn how to use Hugging Face to train the models they were building. We showed them how to use the resources, and they ran with it.”
That advisory relationship cut development time by 30 percent. The team used Hugging Face transformers to train document AI models targeting entity resolution and semantic search—the core tasks needed to turn a messy receipt photo into structured, queryable data. This wasn’t a lift-and-shift migration. Fetch built a shadow pipeline that reprocessed all 11 million daily receipts in parallel with the old system, running audits at full volume before anything touched production. “The black box was still running the show,” noted lead ML engineer Sam Corzine—a cautious approach that kept the switchover from interrupting the rewards flow.
The broader lesson here isn’t about receipts. It’s about the maturing enterprise path for open-source AI. Hugging Face has spent years building credibility with developers; now it’s monetizing that trust through advisory programs that pair its engineers with in-house teams. Fetch didn’t outsource its AI strategy—it internalized it, then brought in expertise to close specific gaps. That’s a model a lot of enterprises watching their own third-party AI dependencies should study. The 30 percent time savings is nice, but the real win is that Fetch now owns its data pipeline end to end.
