Projects

The Problem

Customer churn is one of the most expensive problems in subscription businesses — it's far cheaper to retain a customer than acquire a new one. This project used IBM's public Telco dataset to identify which customer characteristics and behaviors most strongly predict churn, surfacing actionable segments for a retention team to target.

Approach

Exploratory data analysis in pandas across customer attributes — after cleaning (nulls, categorical encoding, TotalCharges dtype fix), I used Matplotlib and Seaborn to visualize churn rates across contract type, tenure, payment method, and service bundle. The goal was a communicable story, not a black-box model.

Key Findings

• Month-to-month customers churned at 3× the rate of annual subscribers — by far the strongest predictor.
• Churn dropped sharply after 12 months of tenure; early-stage customers are the highest-priority retention target.
• Electronic check payers churned at nearly 2× the rate of auto-pay customers.
• Service bundles reduced churn; premium add-ons without a contract did not.

Impact

Clear prioritization framework: focus retention on month-to-month customers in year one paying by electronic check — 22% of the base, disproportionate share of churn.

The Problem

E-commerce businesses generate enormous transactional data but often lack the query infrastructure to answer strategic questions: who are the most valuable customers, are they being retained, and which products belong together? This project built that analytical layer from scratch.

Approach

Three analysis areas in PostgreSQL: CLV segmentation, cohort-based retention, and product affinity. Every query is documented with a plain-English explanation. No ORMs, no Python wrappers — demonstrating that complex business questions can be answered cleanly in SQL alone.

Key Findings

• Top 20% of customers drove 68% of revenue — stronger Pareto than typical benchmarks.
• Month-1 retention 38%; stabilized at ~22% by month 3, minimal drop-off beyond.
• Highest product affinity pair: 4.2× lift above random — strong bundling candidate.
• Promotion-acquired customers had lower CLV despite higher initial order values.

Impact

Reusable query library — drop in any transactional dataset and the CLV and cohort analyses run with minimal modification.

The Problem

Finance teams in mid-size organizations often track budget vs. actuals in fragmented, manually updated spreadsheets that are slow to reconcile and easy to break. This project built a clean, dynamic Excel model that automates variance calculations and surfaces issues at a glance.

Approach

Strict separation of data input, calculation, and presentation layers. Power Query for ingestion; structured tables with named ranges feed a variance calculation sheet; conditional formatting flags over-budget items automatically in the dashboard layer.

Key Features

• Automated variance flags — no manual highlighting.
• Dynamic pivot summary with department and period slicers.
• Rolling year-end forecast from actuals entered to date.

Impact

Demonstrates that well-structured Excel is still a powerful analytical tool in environments where BI platforms aren't available — and that the model is maintainable by anyone, not just its original builder.