Stripe has built a transformer-based payments foundation model that it uses to power fraud detection in Radar. This model is self-supervised and trained on tens of billions of transactions, letting it “discover its own features” rather than relying solely on hand-picked signals. By compressing each payment into a dense, “atomic” embedding, Stripe can spot subtle patterns of card-testing fraud that simpler models miss—and in practice this foundation model boosted card-testing detection rates on large businesses by 64% practically overnight (TechCrunch, Stripe).
Rather than replacing its tiered fraud-detection flywheel, Stripe augments its specialized card-testing models with the transformer’s embeddings. In production, they still run high-level prevalence and mid-level surface-attack models, but then feed individual transactions’ embeddings into sequence classifiers to catch novel or low-signal attacks in real time (Stripe).
Under the hood, that single foundation model isn’t just for fraud. By learning general-purpose vectors for every transaction—much like a language model embeds words—Stripe can fine-tune the same Transformer backbone for multiple tasks: authorization risk, fraud scoring, dispute resolution, and more, rather than maintaining separate models for each use case (LinkedIn).
Beyond fraud, Stripe has also applied Transformer-powered ML to dispute handling. Its Smart Disputes feature uses an AI-powered rules engine to extract and assemble the most relevant evidence from network-wide data, auto-submitting packets on your behalf. Meanwhile, other ML models predict if your account is trending toward excessive dispute activity—letting Stripe proactively alert you before you hit risk thresholds (Stripe Docs, Stripe Docs).
Internally, Stripe’s Applied ML team has pre-trained LLM-style models on over a trillion tokens of Stripe data to build support and documentation assistants, developer help bots, and other AI-driven tools—showing how the same Transformer machinery can power everything from customer support to code generation (gk).