Real-Time Anomaly Detection for Freight Logistics
Capstone project for Radiant Delivers (international freight forwarder), 8 weeks · April 2026 – June 2026
Team: Akanksha Singh, Yuvasri Govindasamy, Shraddha Mahangare, Graeme Tobias Ampeire
Isolation Forest
Extended Isolation Forest
Gaussian Mixture Model
GCP / Vertex AI
Pub/Sub & Cloud Run
ClickHouse
Streamlit
SAP TM
Overview
- Designed and deployed an ML system that flags data-entry errors in freight shipments before they cause downstream cost, running across 1.4M+ SAP Transportation Management documents
- My contribution: data modeling and pipeline. Owned the feature engineering / data modeling that fed the detection models and built the real-time ingestion-to-inference pipeline end to end
- Solved a domain problem, not just a modeling one. Built stage-aware detection that learns when a missing field is actually an error vs. normal workflow (e.g., a completed Road shipment with no delivery date is an error; a completed Sea shipment without one is expected), replacing rigid rule-based logic
- Built three complementary models: an Isolation Forest for structural completeness with stage-aware scoring, an Extended Isolation Forest for pricing anomalies with dollar-impact explanations, and a Gaussian Mixture Model for peer-cluster deviation
- Engineered a real-time GCP pipeline: SAP TM → ClickHouse → Pub/Sub → Cloud Run → Vertex AI → CloudSQL → live Streamlit dashboard
- Shipped a four-page “Data Quality Watchtower” dashboard that converts raw anomaly scores into ranked, plain-language flags the ops team actions every morning
- Delivered a platform evaluation (SAP AI Core vs. Vertex AI) and a 12-month roadmap toward MCP-based agentic integration
- Key insight: domain expertise is what turns a statistically sound model into an operationally trusted tool. Translating “anomaly score: 0.87” into “missing carrier SCAC on a completed linehaul, blocks invoicing” is what gets a dashboard used instead of ignored