Real-Time Fraud Detection & Inference Pipeline
Train and deploy sub-15ms machine learning fraud classifiers on streaming financial transactions.
Problem Statement
Financial platforms must evaluate transaction fraud risks in real-time without introducing checkout latency. You will build a complete ML inference pipeline with feature extraction and ONNX runtime optimization.
What You'll Build
A low-latency fraud scoring microservice that receives transaction payloads, computes feature embeddings, evaluates an XGBoost/PyTorch classifier, and logs drift metrics.
Technology Stack
Skills You'll Master
Requirements & Prerequisites
- Python 3.10+ with PyTorch and scikit-learn installed.
- Basic understanding of classification metrics: Precision, Recall, and ROC-AUC.
Guided Milestones (3)
Step-by-Step ArchitectureData Preprocessing & Feature Pipeline
Handle extreme class imbalance using SMOTE and construct automated feature transformers.
- Feature extraction script
- Baseline model evaluation report
Model Quantization & ONNX Export
Quantize neural weights to INT8 and benchmark latency reduction.
FastAPI Production Inference Server
Serve predictions with p99 latency under 15ms with structured error boundaries.
Project Submission & Verification
Submit your public GitHub repository and live deployment link for automated test checks.
Add to ASCI Portfolio
Proves hands-on competence in production ML model serving and latency constraints.