Fraud Detection: from DataOps to MLOps

💡 Highlights

Data & FeaturesAutoMLOrchestration & InfrastructureExperiment TrackingReal-Time Inference

Highlights

• Built modular, testable SQL pipelines with dbt, enabling reproducible and version-controlled feature generation
• Registered features to Feast (open source feature store) for consistent usage in both batch training and real-time serving
• Enabled feature backfilling and time-travel queries, supporting point-in-time correctness for fraud detection models

Highlights

• Performed distributed Bayesian hyperparameter optimization using Ray Tune + Optuna, accelerating tuning efficiency at scale
• Handled imbalanced datasets using imbalanced-learn to dynamically apply over- and under-sampling strategies, improving model prediction performance
• Ensured reproducibility by tracking fixed random seeds, stratified sampling, and consistent data splits across trials

Highlights

• Deployed the entire pipeline on Kubernetes, enabling scalable, containerized execution of distributed services
• (WIP) Orchestrated pipeline stages with Airflow, improving automation, observability, and task dependency management
• Integrated MinIO (S3-compatible) storage for storing intermediate features and trained models across components

Highlights

• Integrated MLflow to auto-log training parameters, metrics, and artifacts, enabling experiment reproducibility and traceability
• Versioned models and experiments using MLflow’s tracking server, enabling full auditability and rollback
• Stored model artifacts in remote object storage (MinIO), making them accessible for downstream deployment

Highlights

• Deployed models as gRPC and REST endpoints using KServe, supporting diverse integration requirements
• Ensured compatibility between training-time and serving-time features via Feast’s online store integration
• Enabled autoscaling and scale-to-zero, optimizing cost for infrequently used models
• Configured A/B testing traffic split, allowing controlled experimentation in production deployments

---

In real-world machine learning projects, managing the workflow from data transformation to model deployment is often fragmented, error-prone, and hard to scale.

This project demonstrates how to streamline and automate the entire lifecycle—from feature engineering to hyperparameter tuning, model tracking, and deployment—using modern open source tools and running fully on Kubernetes.

The use case for this project is fraud detection, a high-impact and time-sensitive problem where real-time inference is critical. It serves as a practical demo of how to operationalize machine learning pipelines that are version-controlled, reproducible, and ready for production.

It’s designed to be:

• Reproducible – All data transformations, features, and models are versioned via dbt, Feast, and MLflow
• Scalable – Built on Kubernetes, enabling distributed training and resource orchestration across services
• Modular – Each stage is decoupled and replaceable, promoting clear responsibility and reuse
• Open Source – Fully built on open source tools like dbt, Feast, Ray, Optuna, MLflow, and KServe
• Portable – Easily adapted to other use cases beyond fraud detection

Whether you're a data engineer, ML practitioner, or platform builder, this project offers a clear, working example of how to bridge DataOps and MLOps on a scalable, production-ready foundation.

🏗️ Architecture

Architecture (Click to Enlarge)

dbtSQLMeshFeastAirflowRayMLflowKServe

Highlights

• Developed incremental models to process data in minibatches, improving pipeline efficiency and reducing compute cost
• Implemented test coverage and schema validation, ensuring data quality across transformations
• Generated documentation automatically from dbt models, enhancing maintainability and team collaboration

Highlights

Work In Progress

Highlights

• Materialized online features to Redis, enabling real-time feature retrieval for low-latency inference
• Supported both batch and online inference by separating offline and online stores
• Enabled time-travel and point-in-time feature retrieval, ensuring training-serving consistency for fraud detection

Highlights

Work In Progress

Highlights

• Performed distributed Bayesian hyperparameter tuning using Ray Tune and Optuna, accelerating model search and training time
• Integrated imbalanced-learn to automatically select appropriate over- and under-sampling strategies, improving performance on imbalanced datasets
• Scaled training across nodes on Kubernetes, leveraging Ray cluster for efficient resource utilization

Highlights

• Integrated MLflow to auto-log parameters, metrics, and artifacts during training, enabling experiment tracking and auditability
• Logged final model as a versioned artifact, facilitating reproducible deployment and rollback
• Enabled reproducibility across environments by centralizing tracking and storage in a MinIO-based S3-compatible backend

Highlights

• Deployed models as gRPC and REST endpoints using KServe, supporting diverse integration requirements
• Ensured compatibility between training-time and serving-time features via Feast’s online store integration
• Enabled autoscaling and scale-to-zero, optimizing cost for infrequently used models
• Configured A/B testing traffic split, allowing controlled experimentation in production deployments

🗂️ What's Inside?

.
├── dbt/       - Transform raw data into feature tables
├── sqlmesh/   - (Work In Progress)
├── feast/     - Define and manage features with Feast
├── airflow/   - (Work In Progress)
├── ray/       - Run distributed hyperparameter tuning with Ray and Optuna
├── mlflow/    - Track experiments and log models with MLflow
├── kserve/    - Deploy trained models using KServe
├── minio/     - Configure MinIO (S3-compatible) for model/data storage