Client: FinTech App with 50M+ users
Tools: Amazon SageMaker, Apache Kafka, PyTorch, XGBoost
Challenge:
Rising chargebacks and undetected fraud due to rule-based systems not scaling with user base.
Solution:
- Built a real-time fraud detection engine using SageMaker and PyTorch
- Ingested transaction data via Kafka streams
- Trained ensemble models (XGBoost + deep neural nets) on behavior patterns
- Integrated with transaction approval APIs and alerting systems
Outcome:
Flagged 97% of fraudulent transactions under 2 seconds. Reduced financial fraud losses by 40%.