mansoorsk/procurement-ltv

Procurement Lead Time Variance — ML System

An end-to-end machine learning system to predict delivery variance for Olist e-commerce orders.

Project Overview

Predicts how many days early or late an order will be delivered relative to the estimated delivery date using the Brazilian Olist e-commerce dataset.

Model

• —Algorithm: XGBoost (Optuna-tuned, 50 trials)
• —Test MAE: 4.58 days
• —Test RMSE: 7.41 days
• —Directional Accuracy: 91.67%
• —Prediction Intervals: Conformal prediction (90% coverage)

Features Used

44 engineered features including:

• —Estimated delivery window (strongest predictor)
• —Seller historical performance
• —Geographic features (same state / region)
• —Seasonal features (purchase month, day of week)
• —Product dimensions and weight
• —Payment features

App Structure

• —User Tab — Single order predictor with confidence intervals
• —Developer Tab — MLflow experiment tracking + Evidently AI drift monitoring + Model performance

Tech Stack

• —Python, XGBoost, LightGBM, Scikit-learn
• —MLflow (experiment tracking)
• —Evidently AI (drift monitoring)
• —SHAP (model interpretation)
• —Streamlit (frontend)
• —Docker (containerization)
• —GitHub Actions (CI/CD)
• —Hugging Face Spaces (deployment)

Pipeline Phases

1. 1.Problem Definition
2. 2.Data Collection (Supabase + SQLAlchemy)
3. 3.Target Construction + Train/Val/Test Split
4. 4.EDA (train split only — no leakage)
5. 5.Data Cleaning
6. 6.Feature Engineering
7. 7.Model Training (Baseline → Linear → RF → XGBoost → LightGBM)
8. 8.Hyperparameter Tuning (Optuna)
9. 9.Model Validation + Conformal Prediction Intervals
10. 10.Model Interpretation (SHAP)
11. 11.Deployment (Streamlit + Docker + GitHub Actions + Hugging Face)
12. 12.Monitoring (Evidently AI)