Balancing LLM-driven automation with human oversight at Stripe requires a "human-in-the-loop" architecture. My approach focuses on using AI for high-velocity signal synthesis, case triage, and evidence summarization to drastically reduce investigation time. By surfacing high-confidence patterns and automating routine data collection, we empower expert reviewers to focus on high-judgment, complex decisions. I design these systems to be explainable, providing investigators with clear "reasoning trails" alongside AI suggestions. This minimizes the risk of hallucinations while maintaining strict regulatory compliance. Ultimately, I build tools where the LLM acts as an augmented intelligence partner, ensuring human expertise remains the final authority on critical financial crime outcomes.

To measure the success of a multi-tenant risk platform like Stripe’s, we focus on three primary dimensions: Performance, Efficacy, and Developer Experience.

1. System Reliability & Latency: We track p99 latency and uptime, ensuring the screening engine meets strict SLAs without impacting transaction throughput.
2. Risk Detection Efficacy: We measure the precision and recall of our matching and policy engines, monitoring false positive rates to ensure investigators remain efficient.
3. Platform Adoption & Velocity: We track "time-to-onboarding" for new product teams using the multi-tenant APIs, measuring the ease of integrating new policy logic as code.
4. Operational Efficiency: We quantify the reduction in manual review time per case through successful AI-augmented tooling and workflow automation.

To ensure policy as code remains scalable as global regulations evolve, I would architect a decoupled, multi-tenant policy engine that separates policy definition from enforcement. By implementing a domain-specific language (DSL) for representing complex sanctions and AML logic, we can allow Compliance teams to update rules rapidly without requiring core infrastructure deployments. I would prioritize an asynchronous evaluation pipeline that abstracts policy logic from transaction flows, utilizing caching and intelligent indexing to maintain strict latency SLAs. Furthermore, I would build automated testing and observability suites to validate policy changes against historical traffic, ensuring that as global regulations shift, our platform provides consistent, auditable, and performant enforcement across all Stripe products.

The Financial Crimes Engineering team is re-architecting its screening platform to enable seamless, scalable integration for new Stripe products and subsidiaries. Their strategy focuses on three primary objectives: implementing a multi-tenant Risk Platform, transitioning to a "policy as code" model, and enhancing system architecture to handle global regulatory requirements.

By evolving the infrastructure into a multi-tenant offering, the team allows internal products and external developers to utilize Stripe’s screening capabilities via APIs. This modernization focuses on balancing low-latency performance with high-accuracy detection. Ultimately, this structural shift ensures that new products launch with robust, compliant controls, enabling the company to express richer, automated risk policies while maintaining strict uptime and SLA guarantees.

The Financial Crimes team collaborates with ML experts to reduce false positives by integrating AI-augmented tooling directly into their investigation and screening workflows. They translate complex regulatory requirements into actionable software, utilizing ML models to automate signal acquisition and synthesis. By building infrastructure that feeds into human-in-the-loop review systems, the team ensures that ML models assist expert reviewers in making faster, more accurate case resolutions. Furthermore, engineers work closely with ML teams to apply technologies like entity resolution, fuzzy string matching, and NLP to improve detection pipelines. This synergy allows the team to balance screening latency with precision, ensuring that the system identifies true bad actors while minimizing friction for legitimate business users.