Messy data.

Entry in the Codebasics Resume Project Challenge: a data-driven analysis of a business crisis.

• Analyzed 6 connected datasets (orders, customers, deliveries, ratings, restaurants) with Pandas, Seaborn and Matplotlib.
• Compared performance before and during the crisis.
• Found a 69% order decline, a 71% revenue drop, and SLA compliance falling from 43.6% to 12.2%.

View on GitHub

SQL analysis of a five-table relational sales database in PostgreSQL.

• Seven modules covering joins, aggregations, subqueries and window functions.
• Answered customer lifetime value, sales rep performance and marketing channel attribution using RANK(), NTILE(), LEAD()/LAG() and CTEs.
• Organized as modular SQL files by topic, with an ERD mapping every table relationship.

View on GitHub

An ML system for finance teams that predicts freight costs and flags risky invoices before they are paid.

• Built an end-to-end ML system on vendor invoice and purchase data (SQLite, 5,543 records) to predict freight costs and flag high-risk invoices for manual review.
• Trained and compared linear regression, decision tree, and random forest regressors for freight cost prediction and a random forest classifier for invoice risk flagging, achieving R² = 0.97 (freight) and 89% accuracy (risk flagging) after GridSearchCV tuning.
• Deployed both models via an interactive Streamlit application for real-time freight prediction and invoice risk scoring.

View on GitHub

Survey analysis of what drives airline passenger satisfaction, plus a model to predict it.

• Analyzed 103,904 airline passenger survey records across 23 features using Pandas, Seaborn and Matplotlib, cleaning the data (median imputation for missing arrival delay) and engineering age and flight-distance groups.
• Found that business travelers in Business class were 72% satisfied, versus only 9 to 12% of personal travelers in every class, and that higher seat comfort, entertainment and online boarding ratings went with higher satisfaction.
• Compared Decision Tree (93.8%) and XGBoost (95.9%) classifiers against a 56.7% baseline, dropping highly correlated features to avoid data leakage and tuning with GridSearchCV.

View on GitHub