Latest Project · May 2026
Olist RFM Customer Segmentation: Brazilian E-Commerce
PythonPandasNumPyScikit-learn (KMeans)Matplotlib
Raw orders→
RFM scoring→
11 segments→
KMeans (k=4)→
Opportunity sizing
Segmented 93,357 customers (R$ 15.4M revenue, 2016 to 2018) with RFM in Python, running two scoring methods in parallel. Caught a bug in the reference tutorial that aggregated by customer_id instead of customer_unique_id, which had collapsed Frequency to 1 for every customer, and fixed it. Quantified R$ 1.88M of addressable opportunity, about 12% of revenue, across 11 segments. The largest single play: convert 7,232 Potential Loyalists into repeat buyers inside the 30-day window, worth R$ 634K, validated by KMeans (k=4) on log-RFM.
View on GitHub →
AI-Assisted BI · Apr 2026
AdventureWorks DTC Customer Profitability
Claude AIPower BI Modeling MCPDAXSQL Server
SQL Server→
Power BI model (MCP)→
22 DAX measures→
Dashboards
Built a customer profitability model on AdventureWorksDW2022 (60,398 DTC rows, 18,484 customers). Used Claude through the Power BI Modeling MCP to write 22 DAX measures and 5 calculated columns in the live model, and reviewed every expression. Caught a reserved-word collision and a heavy calculated column before they shipped. Found the top 20% of customers drove 66.3% of revenue, and mid-income $40 to 100K customers drove 62% of profit ($7.5M of $12M).
View on GitHub →
Nov 2025
S&P 500 Financial Performance: Post-Pandemic Recovery
SQL ServerPower BIDAXR (ARIMA)
SQL Server→
DAX + R (ARIMA)→
6 dashboards
Financial-statement analysis for about 500 S&P 500 companies in SQL Server and Power BI, with 13 DAX measures. Quantified aggregate revenue growth from $5.6T to $13.1T and net income margin from 8% to 11%, with full P&L decomposition and YoY variance across 6 dashboards. Built an ARIMA model in R, connected to Power BI, to project 4 quarters of forward revenue with confidence bands.
View on GitHub →
Oct 2025
AdventureWorks: U.S. Sales & Profitability
SQL ServerPower BIDAXPower Query
U.S.-focused profitability model from AdventureWorksDW2022 (33,400 USD transactions) with 20 DAX measures: Revenue YTD, Profit Margin %, Pareto ranking, and 3-month rolling averages, across 3 dashboards. Found Western regions drove revenue through volume while Eastern regions delivered higher margins, and Accessories ran 62.6% margin vs 41 to 45% for Bikes.
View on GitHub →
Mar 2026
BF LUBS: Revenue/Volume Variance Diagnostic
PythonPandasNumPyMatplotlib
Diagnosed a paradox in 601,836 lubricant transactions (2018 to 2024) where revenue grew while volume fell. Used IQR outlier detection with an operational filter (excluded values above the 20,000L vehicle capacity) plus Pareto analysis, and showed one region-market-customer combination drove 28.95% of total volume loss. Wrote 5 retention recommendations naming specific accounts and the flagship product.
View on GitHub →
Apr 2026
Customer Retention Analytics: B2B Distributor
PythonScikit-learnPandas
Churn risk-scoring system from 11,233 sales records and 544 customer profiles. Trained a Random Forest classifier and scored every active 2025 account by risk tier. Found days-since-last-purchase was the strongest signal (about 20% feature importance) and month-to-month customers churn at 2x the rate of annual contracts. Recommended a 60-day silence flag and contract-tier incentives.
View on GitHub →
Mar 2026
BF LUBS: Logistics Cost-Impact Analysis
PythonPandasFoliumOpenRouteService API
Delivery data→
OpenRouteService API→
Route simulation→
Cost impact
Route-simulation tool that called the OpenRouteService API across 6 delivery days. Quantified that an emergency branch change added 4,728 km to the weekly route (+85.6%, from 5,522 km to 10,250 km). Recommended tighter inventory control and pre-built contingency routes.
View on GitHub →
Mar 2026
Customer Geographic Segmentation & Auto-Classification
PythonScikit-learnFolium
Segmentation pipeline combining unsupervised KMeans (5 geographic divisions chosen by the elbow method) with a supervised Decision Tree, so new customers route to the correct commercial division automatically.
View on GitHub →