Project Title: AI-Driven Supply Chain Risk & Delay Predictor A Full-Stack Data Science Solution utilizing Medallion Architecture, Transformers (FinBERT), and XGBoost to predict logistics disruptions.
Executive Summary This project addresses the "Predictive Gap" in global supply chains by merging internal logistics records (180,000+ rows) with external unstructured data (11,000+ news headlines). By building a SQL-based Medallion Architecture, I transformed messy CSV data into a validated "Gold Layer," enabling an XGBoost model to predict shipping delays with a Mean Absolute Error (MAE) of <1 day.
The Architecture (Medallion Pattern) To ensure data integrity and scalability, the pipeline is organized into three distinct tiers:
Bronze Layer (Raw): Immutable landing zone for raw DataCo CSVs and JSON news datasets.
Silver Layer (Cleaned): Implementation of a SQL Quality Firewall that identified and filtered out ~2% of corrupted records (e.g., negative shipping times).
Gold Layer (Model-Ready): A high-performance feature set created by performing a temporal join between logistics dates and aggregated daily news sentiment.
graph TD
A[Raw DataCo CSVs] -->|Ingestion| B(Bronze Layer: Raw SQL)
B -->|SQL Quality Firewall| C(Silver Layer: Cleaned)
D[News Headlines] -->|FinBERT NLP| E(Sentiment Risk Scores)
C -->|Temporal Join| F(Gold Layer: Model-Ready)
E -->|Temporal Join| F
F -->|Train/Test Split| G{XGBoost Predictor}
G -->|Deployment| H[Streamlit Dashboard]
Business Value & ROI Operational Efficiency: By identifying high-risk delays 48 hours in advance, logistics managers can proactively reroute shipments, saving an average of $500 per shipment in expedited freight costs.
Risk Mitigation: The inclusion of NLP-driven news sentiment identifies market-wide "Risk Multipliers" that traditional logistics data alone misses.
The NLP Engine (Alternative Data) A primary innovation of this project is the inclusion of "Alternative Data" to capture external market risks:
Model: Utilized FinBERT (a specialized Financial Transformer) to perform sentiment analysis on global headlines from 2015–2018.
Impact: The engineered daily_risk_score feature ranked in the Top 25% of predictors, successfully capturing how geopolitical and economic news directly impacts regional shipping performance.
Model Performance & Insights I compared multiple ensemble methods to find the optimal balance between stability and accuracy:
Random Forest: Provided a stable baseline (MAE: 0.98 days).
XGBoost: Selected for deployment due to high performance on high-cardinality categorical data (MAE: 0.99 days).
Key Discovery: While "Shipping Mode" remains the strongest predictor, News Sentiment provided a critical "Risk Multiplier," accurately flagging delays during periods of high market volatility.
Tech Stack Data Engineering: Python, SQLite, Pandas.
Machine Learning: XGBoost, Scikit-Learn.
NLP: HuggingFace Transformers (FinBERT), PyTorch.
Deployment: Streamlit, Joblib.
How to Run Locally Clone: git clone https://github.com/karllennon/Supply-Chain-Risk-Predictor-.git
Environment: python3 -m venv venv && source venv/bin/activate
Install: pip install -r requirements.txt
Launch Dashboard: streamlit run scripts/app.py