This Vietnamese fintech startup had built a mobile-first digital wallet and payments platform serving over 6 million users, processing approximately VND 42 trillion (roughly USD 1.7 billion) in transactions monthly. After raising a USD 22 million Series B, the company was scaling rapidly — adding 400,000 new users per month — but fraud was scaling even faster.
Fraud losses had reached 0.34% of transaction volume, nearly double the industry benchmark of 0.18%. At the company's transaction volume, this represented approximately USD 5.8 million in annual losses. The fraud team of eight analysts was overwhelmed, manually reviewing 12,000 flagged transactions daily using rule-based triggers that generated a false positive rate of 89% — meaning nearly 9 out of 10 flagged transactions were legitimate, wasting analyst time and creating friction for honest users.
The fintech's rapid growth made it a target for sophisticated fraud networks using techniques including synthetic identity creation, social engineering, and organized account takeover rings. The State Bank of Vietnam was also increasing scrutiny of fintech fraud controls, and the company's banking partners had threatened to revoke partnership agreements if fraud rates were not brought under control within six months.
Pertama Partners' AI Readiness Audit analyzed the fintech's transaction data, user behavioral patterns, device fingerprinting data, and historical fraud cases. We found that the existing rule-based system was flagging transactions based on simplistic thresholds — amount, frequency, and recipient — that failed to distinguish between legitimate power users and genuine fraudsters. The audit also identified that 73% of confirmed fraud followed identifiable patterns detectable with machine learning.
Our AI Pilot Program deployed a multi-layered fraud detection system. The first layer performed real-time transaction scoring using a gradient-boosted model that analyzed over 150 features including transaction context, device behavior, network analysis, and temporal patterns. The second layer used graph neural networks to detect organized fraud rings by analyzing transaction networks and identifying suspicious clustering patterns. The third layer performed continuous user behavior profiling using an anomaly detection model that learned each user's normal pattern and flagged deviations.
The system was engineered for sub-300-millisecond scoring to maintain the seamless payment experience users expected. AI Governance Retainer engagement established model monitoring, fairness testing (ensuring the system did not disproportionately flag users based on demographic factors), and regulatory reporting aligned with SBV requirements. Team Training equipped fraud analysts to investigate AI-flagged cases efficiently and provide feedback that continuously improved model accuracy.
"Our previous fraud system was like a smoke detector that went off every time someone cooked dinner. Pertama Partners built us a system that only sounds the alarm when there is actually a fire — and it catches fires we never could have detected before."— Tran Duc Anh, Chief Technology Officer
Every transformation starts with a conversation. Let's discuss your challenges and opportunities.
Discuss Your Challenge