PayGuard
Real-time payment-fraud detection on a streaming medallion pipeline.
Problem
Payment fraud must be caught in the moment a transaction clears, not in a nightly batch. Latency directly translates to loss, and models must score a continuous stream without falling behind.
Solution
A streaming pipeline scores transactions as they arrive, refining raw events through a medallion architecture and serving a gradient-boosted model behind a low-latency API with an interactive demo.
Architecture
Kafka streams transactions into Spark Structured Streaming, which writes bronze, silver, and gold layers to Delta Lake. A LightGBM model is trained on the curated features and served through FastAPI, with a Streamlit app for live scoring and exploration.
Tech stack
KafkaSpark Structured StreamingDelta LakeLightGBMFastAPIStreamlit