Supply Chain Disruption Analytics

Solo project · 2026 · deployed

#data-engineering#cloud#bi#machine-learning#python

Key Result

A full analytics lifecycle — ingestion, monitoring, statistical validation, and predictive modeling — translated into a capacity-based prioritization policy, not just a trained model.

Overview

Shipment delays directly affect inventory availability, supplier evaluation, and revenue — and in many organizations, delays are only detected after they’ve already happened. This project builds a layered supply chain analytics system on Azure that creates reliable visibility into shipment performance, validates delay drivers statistically, and then moves toward order-time delay-risk prediction.

The objective was not just to train a model, but to design a complete analytics lifecycle: data ingestion → data engineering → monitoring → statistical validation → predictive modeling → operational policy.

Architecture

Raw CSV
  -> Azure Blob Storage (raw layer)
  -> Azure Data Factory transformation
  -> Curated Parquet dataset
  -> Statistical validation (delay drivers)
  -> Predictive delay-risk model
  -> Power BI dashboard

Design decisions worth calling out: raw data is preserved to maintain lineage and reproducibility; the curated Parquet layer exists for performance and a standardized schema; transformation runs in Azure Data Factory specifically to demonstrate cloud-native ELT rather than local scripting; and compute is kept separate from reporting so analytics logic isn’t coupled to the visualization layer.

Results

The statistical validation layer confirmed which delay drivers were real versus noise before any model was trained on them — avoiding the common failure mode of fitting a predictive model to spurious correlations. The resulting risk-scoring model is framed as a ranking tool: it prioritizes which shipments deserve intervention capacity first, rather than claiming to predict delay with certainty. A downloadable Power BI dashboard (.pbix) ships alongside the pipeline for direct inspection.

Code Highlights

Validation and modeling are kept in clearly separated scripts (src/sanity_checks/, src/actionable_insights/) so the statistical case for each delay driver is auditable independently of the final model — a reviewer can check the evidence before trusting the score.

What I'd Improve

The predictive layer is deliberately framed as a ranking tool for prioritizing limited intervention capacity, not a black-box delay predictor; the next step is closing the loop with outcome data on whether the prioritized interventions actually reduced delays.