I work on bringing my interest in data and AI into real business processes.

Step by step

1. 1

   Problem

   Forecast the daily electricity use of 7,036 transformers for the next 122 days, using 1.2 million rows of real data from İzmir and Manisa. Scored with RMSLE.

2. 2

   Data cleaning

   The data was profiled with pandas. 444 physically impossible readings were caught with an installed power times 24 hours rule and corrected.

3. 3

   Feature engineering

   25 features were built from the calendar, location, installed power, each transformer's recent history and temperature data from the Open Meteo API.

4. 4

   Model

   A LightGBM model forecasts the whole period at once instead of day by day, and its output is blended with each transformer's last 30 days.

5. 5

   Result

   Testing one change per submission, the Kaggle score improved from 1.13709 to 1.05008. A transformer's own recent history proved to be the strongest signal.

Step by step

1. 1

   Question

   Which listing features change Airbnb prices in Istanbul? Public Inside Airbnb data: 22,865 listings, 15 variables, prices in Turkish lira, analyzed in R.

2. 2

   Descriptive statistics

   Mean price 2,752 ₺, median 2,412 ₺, skewness 0.91. Most listings sit between 1,000 and 3,000 ₺, with a long tail of luxury listings.

3. 3

   Normality check

   Q-Q plot, Shapiro-Wilk (W = 0.935) and Kolmogorov-Smirnov tests rejected normality. With this sample size, the Central Limit Theorem still supports parametric tests.

4. 4

   Hypothesis tests

   Welch t-test: the European side is about 206 ₺ more expensive (p < 0.001). One way ANOVA with Tukey HSD: entire homes cost about 1,137 ₺ more than private rooms.

5. 5

   Result

   Location and room type both have a significant effect on price, and about 70 percent of listings are entire homes.

Step by step

1. 1

   Problem

   Forecast 15 days of sales for every store and product family, using 3,000,888 daily records plus store, holiday, transaction and oil price tables. Scored with RMSLE.

2. 2

   Exploratory analysis

   EDA with pandas and Seaborn: steady growth, a strong weekly rhythm, and 31.3 percent of rows with zero sales.

3. 3

   Feature engineering

   31 features from the calendar, holidays, store metadata, transactions and oil price, since Ecuador's economy depends on oil.

4. 4

   Model

   LightGBM on a log transformed target with a chronological split: trained up to late July 2017, the last 15 days held out for validation.

5. 5

   Result

   Validation RMSLE of 0.3691. Store type is one of the strongest signals: type A stores sell about 3.5 times more than type C.

Step by step

1. 1

   Business question

   HR planned to reward Branch B's manager (77.4 percent satisfaction) and replace Branch A's (53.9 percent). Was the gap real?

2. 2

   Data preparation

   Survey results were queried with SQL and analyzed with pandas, split by branch and by location type: mall or industrial site.

3. 3

   Hidden factor

   80 percent of Branch A's surveys came from industrial sites, where satisfaction is naturally lower. For Branch B it was only 20 percent.

4. 4

   Stratified analysis

   Within each location type Branch A scores higher. Two proportion z tests: significant in malls (p = 0.017), no difference in industrial sites (p = 0.566).

5. 5

   Recommendation

   The report advised HR not to act on the aggregate rate: the original decision would have penalized the better manager. A classic case of Simpson's paradox.

Step by step

1. 1

   Data

   Olist's public marketplace data from 2016 to 2018: 99,441 orders, 112,650 items and 96,096 customers across linked tables.

2. 2

   SQL analysis

   Tables were joined and queried in SQLite to answer five business questions: revenue sources, growth, geography, delivery and repeat purchases.

3. 3

   Findings

   The top 5 categories bring 39.7 percent of revenue, Black Friday 2017 was the peak day with 1,166 orders, and São Paulo holds 42 percent of customers.

4. 4

   Operations

   Delivery takes 12.6 days on average and 8.11 percent of orders arrive late. Only 3.12 percent of customers order again, a clear growth area.

5. 5

   Dashboard

   A two page Power BI report with KPI cards, revenue by category, a daily trend and a state by category matrix.

Step by step

1. 1

   Problem

   Binary classification: predict whether a customer buys, yes or no, from 8 features such as age, income, product price and discount.

2. 2

   Data and split

   1,000 synthetic observations generated with a fixed seed, 72.9 percent positive. 80/20 train and test split.

3. 3

   Models

   KNN with scaling fit on training data only, a CART decision tree (rpart) tuned on depth, and a radial SVM (e1071) tuned on cost.

4. 4

   Results

   Test accuracy: KNN 73.5 percent (k = 27), SVM 72.5 percent (cost = 1), decision tree 72.0 percent (depth 3).

5. 5

   Interpretation

   All three stay close to the 72 percent majority baseline, a reminder that accuracy alone is misleading on imbalanced data.

Step by step

1. 1

   Student performance analysis

   Exploratory analysis with NumPy and pandas, group wise imputation of missing values and correlation analysis.

2. 2

   Learning analytics pipeline

   Cleaning and feature engineering, then logistic regression to flag at risk students and K-Means with PCA to group them.

3. 3

   Smart campus study

   Weekly LMS activity, study hours and final scores were combined to segment students by engagement and performance.

Step by step

1. 1

   Goal

   Find the busiest pickup zones and time slots in very large trip files, fast and without heavy frameworks.

2. 2

   Ingestion

   Streams input with optimized C++ I/O and skips malformed rows instead of stopping the run.

3. 3

   Aggregation

   A composite hash key (zone and hour) counts every slot in constant average time, with memory reserved up front.

4. 4

   Top K and tests

   partial_sort returns the top results in O(N log K). A Catch2 suite of nine tests covers broken data, edge hours, ordering and speed.

Step by step

1. 1

   Data

   500 synthetic hospital cases with specialty, insurance, length of stay, case severity (CMI) and revenue.

2. 2

   Revenue drivers

   Regression shows revenue rising with length of stay and case severity, split by payer type.

3. 3

   Bed efficiency

   Revenue per bed day compares departments: pulmonology and general surgery lead, internal and family medicine trail.

4. 4

   Segmentation

   K-Means on revenue, length of stay and severity, with the number of clusters chosen by the elbow method, and residual based anomaly flags.