Fraud Detection
Reimagined
How I shifted the platform's trust infrastructure from reactive rule-following to layered risk intelligence, cutting fraud and false positives at the same time.
Owned the fraud and trust surface at a B2B sales intelligence platform. Identified that existing rule-based systems were both reactive and blunt, flagging legitimate users while missing sophisticated bad actors. Led a behavioral data analysis to quantify signal value, then designed a two-phase approach: heuristic rules for immediate impact, followed by an in-house AI model with a feedback loop for continuous learning.
A platform scaling fast, with a fraud problem growing faster.
A B2B sales intelligence platform used by tens of thousands of sales professionals to find and enrich contact data. As the platform scaled, a more sophisticated fraud problem emerged: bad actors exploiting free-tier access, scraping data at scale, and evading detection through fragmented rule-based systems that flagged the wrong users and missed the right ones.
I owned the fraud and trust surface as PM, working with a data science team, a security engineer, and cross-functional stakeholders across legal, customer success, and support.
Catching the wrong people, and missing the right ones.
A strategic pivot to expand free trials created an unexpected vulnerability. Our third-party provider analyzed domains, IPs, and signup signals, but had no visibility into what users actually did inside the platform. In Q3 2024 alone, we manually blocked 5,390 fraudulent accounts.
The existing fraud system had two compounding failure modes: it was reactive (rules added after fraud was discovered) and blunt (high false positives flagging legitimate enterprise customers).
Tightening fraud controls to catch more bad actors while simultaneously reducing friction and false positives for legitimate users. Two goals that felt contradictory until the underlying issue was clear.
The data we already had was the advantage no external tool could replicate.
Fraudulent accounts had clear in-platform patterns: rapid account creation under the same domain, phone credit consumption within seconds of verification, multiple signups from the same visitor ID within hours. I led a structured analysis to quantify exactly how many fraudulent accounts each behavioral signal would have caught, giving me both the evidence and the prioritization framework I needed.
Deploy behavioral rules to catch the clearest fraud patterns immediately. Four automated suspension rules running twice daily via Kafka, with validation logic ensuring no paid accounts were ever affected.
Build a self-learning model with a feedback loop: support resolutions and CS override decisions fed back into training data. The system improves quarter-over-quarter without requiring manual rule updates.
Fraud work touches every team. Alignment was the product.
Legal had requirements around automatic versus human-review actions. Customer success was protective of enterprise accounts. Support needed visibility into why accounts were flagged to resolve tickets effectively.
I mapped each stakeholder's constraint, identified where they were in genuine conflict versus just uninformed about how the system worked, and built escalation tiers that accommodated the real constraints. Enterprise accounts got a higher intervention threshold and direct CS notification rather than automated blocking.
The results that arrived together proved the premise.
The 15% reduction in undetected fraud and 20% drop in false-positive tickets arrived together. The point of the redesign was that these weren't in opposition. Catching fraud more accurately meant fewer good users caught in the crossfire.
Fraud is an adversarial problem: bad actors adapt. A system that learns from its own decisions adapts faster than one that waits for a PM to write a new rule.
"Fraud is an adversarial problem. A system that learns from its own decisions adapts faster than one that waits for a PM to write a new rule."