How Chipper Cash Uses Pinecone Vector Search to Stop Fraud in Real-Time

Chipper Cash is a fintech company on a mission to expand financial access across Africa, offering seamless cross-border money transfers to more than five million customers in seven countries. As the platform grew, so did its use of new-user promotions — incentive campaigns designed to attract legitimate customers and drive transaction volume. These promotions, however, also attracted bad actors looking to exploit the system by creating multiple fake accounts to repeatedly claim new-user rewards.

To combat fraud, Chipper Cash initially relied on a third-party identity verification service that performed KYC checks, including validation of government-issued IDs and detection of duplicate accounts. While effective at smaller scale, the solution became a serious bottleneck as user volumes surged — particularly during promotional periods. Verification delays stretched to as long as 20 minutes, giving fraudsters a window to redeem rewards before being caught. Over a six-month period, 16% of the company's promo budget was lost to fraudulent sign-ups, and several campaigns had to be shut down early, limiting legitimate user acquisition.

Samee Zahid, Director of Engineering at Chipper Cash, led the effort to build a faster, AI-native alternative. The core idea was to use visual similarity search: convert each new user's selfie into a vector embedding using a convolutional neural network (ConvNet), then query a vector database to find near-identical embeddings from existing users. A high similarity score would flag a likely duplicate account in real time. The team validated the concept with an open-source vector database, but quickly found that managing it at production scale introduced unacceptable operational overhead. After evaluating managed alternatives, they selected Pinecone for its ultra-low latency, scalability to billions of vectors, and minimal infrastructure burden.

The new system — built and deployed in under a month — integrates Pinecone with Snowflake as the data warehouse and a proprietary Facial Similarity Service (FSS) that generates embeddings and orchestrates queries. When a new user submits a selfie, FSS embeds the image and queries Pinecone, which returns the top three closest matches in under 200 milliseconds. The Chipper backend then evaluates match likelihood alongside relevant metadata to make a final determination. End-to-end, the entire verification process now completes in under 2 seconds.

Since going live, Chipper Cash has achieved 95%+ selfie verification accuracy, reduced fraudulent sign-ups by 10x across all markets, and reclaimed promo budget that now flows exclusively to legitimate new customers. Legitimate users enjoy a frictionless onboarding experience, and the company can run longer, more effective promotional campaigns. Encouraged by these results, the engineering team is actively exploring additional AI applications powered by Pinecone.