Anomaly Detection for Supply Chain Operations

The challenge

Supply-chain anomalies were spotted late, often through manual checks, and each incident carried fulfilment and financial risk. Labels were sparse and noisy, so a fully supervised classifier was impossible. The goal was to build a system that could surface irregularities early, suggest a probable root-cause class, and integrate into existing operations without creating a new silo.

Approach

• Designed a weakly-supervised ensemble: a variational autoencoder for representation learning and density estimation, combined with RQ-Kmeans/RQ-GMM clustering and XGBClassifier/QDA classification heads to score and classify anomalies.
• Built data preprocessing pipelines on AWS Glue and Lambda, with schema validation, drift checks, and lineage tracking so upstream changes do not silently break model inputs.
• Trained models on Amazon SageMaker with quantization and automatic model-registry writes to MLflow, then promoted only versions that passed CI tests and benchmark thresholds.
• Provisioned infrastructure as code with Terraform and Docker, spanning training, API and front-end services across EC2 and Kubernetes with environment parity between dev, staging and production.
• Exposed predictions through FastAPI services and a Streamlit review front-end, with Prometheus/Grafana monitoring for latency, throughput and data-drift alerts.
• Automated CI/CD with GitHub Actions, including unit tests, integration tests, packaging and canary deployments.

Outcome

The platform combined four complementary model families into a single anomaly scoring and classification service. End-to-end MLOps pipelines reduced model update time from weeks to days, and the project was estimated to deliver over 1.6 million Euros in savings by detecting and resolving anomalies earlier than the previous manual process.