Marketplaces · Experimentation, Forecasting & Applied ML, Data Platforms & Reliability

Fraud detection uplift

Dating Marketplace

Rebuilt trust & safety analytics and launched fraud models that cut abusive users from 10% to 0.6%.

Problem

Large volumes of short lived accounts made manual review ineffective. Losses, user trust, and compliance targets were at risk.

Solution

  • Data pipeline overhaul (dbt + Airflow + ClickHouse) to unify trust & safety signals
  • Python/Flask microservices with retraining workflow, champion/challenger evaluation, and governance
  • Shadow evaluation, playbooks, and enablement to launch safely across regions
  • Established daily standups with product, security, and leadership to keep scope tight and surface regulatory feedback.

Outcome

  • Significant uplift in recall at the target precision alongside clear business guardrails
  • ~40 days earlier average detection of bad actors with automated actioning
  • Weekly exec reporting and clear SLOs for pipeline health and model drift