Starling cuts marketplace fraud exposure with Gemini-powered Scam Intelligence tool

Digital neo-banks like Starling operate in an environment where customers transact on marketplaces — platforms such as Facebook Marketplace. Starling needed an intervention that could assess scam risk at the moment of decision and educate customers on warning signs.

Starling built its Scam Intelligence tool on Gemini's multimodal large language model capabilities, deployed through Google Cloud's Vertex AI infrastructure. The approach centers on image-based ad analysis: customers submit a photo of a marketplace listing and Gemini processes the image, flagging red flags such as suspiciously low prices, seller pressure tactics, or other hallmarks of fraudulent listings. The tool returns a risk assessment that helps customers decide whether to proceed — simultaneously educating them on warning signs for future transactions. The Gemini prompt is continuously evaluated and managed through Vertex AI Pipelines, ensuring fraud intelligence stays current as scam patterns evolve. Model Armor guardrails constrain the tool strictly to its intended scope. Development used Vertex AI Workbench with ML pipelines that let the training code run without manual intervention.

The first phase of Scam Intelligence reached consumer testing in just four weeks — a timeline made possible by codifying the entire model training pipeline from day one and letting training code run without manual intervention. Key outcomes:

• 300% increase in the rate at which customers cancel marketplace payments after reviewing Scam Intelligence assessments
• 4-week build cycle from development to first consumer testing phase, demonstrating the delivery velocity achievable with Vertex AI Workbench and automated ML pipelines

Beyond the headline numbers, the tool places risk assessment in the customer's hands at the moment of decision.

• Gemini's multimodal capability enables image-based ad analysis — a fraud-prevention approach impossible with text-only models and directly relevant to peer-to-peer marketplace scams.
• Codifying the full model training pipeline upfront, not as a post-launch retrofit, enabled the four-week delivery timeline.
• Model Armor guardrails are essential when deploying generative AI in regulated financial services — they define the boundary between a useful customer tool and an institutional liability.
• Educating customers on warning signs rather than silently blocking transactions builds durable fraud resistance that compounds beyond the tool itself.