Data Scientist focused on Deep Learning and Computer Vision (CNN). I build end-to-end ML pipelines with emphasis on model reliability, interpretability, and performance. Experienced in working with structured and semi-structured data, applying statistical thinking and data-driven approaches to solve real-world problems. Combine analytical rigor with practical ML implementation, focusing on model reliability, interpretability, and performance improvement.

🔗 LinkedIn: https://www.linkedin.com/in/yuliabezginova/
📩 Email: ybezginova.rs@gmail.com
📨 Telegram: https://t.me/ybezginova_de

Objective: To build a binary classifier distinguishing real from AI-generated (StyleGAN) face images using custom CNN architectures with frequency-domain preprocessing.

Approach: Five sequential experiments — from a baseline CNN to a final architecture with residual blocks, SE (Squeeze-and-Excitation) attention, FFT and HighFreq channels, and dual Average+Max pooling. Class imbalance was addressed via offline augmentation and WeightedRandomSampler.

Best Result: Validation F1 = 0.9835 (threshold = 0.31), submitted to Kaggle. Key techniques: ResidualBlock · SEBlock · AddFFTChannel · AddHighFreqChannel · MedianDenoise · offline minority class augmentation. Code: Notebook 5 (final, Kaggle submission)

Objective: To train a model that classifies user messages as either complaints or questions. Data: Complaints and Questions. Code: Jupyter Notebook

Objective: Build a CNN with skip-connections for image classification on Fashion MNIST. Result: test_acc ~93% after tuning learning rate, batch size, and architecture. Stack: Python, PyTorch Code: Jupyter Notebook

Objective: Train a standard and variational autoencoder (VAE) on CIFAR-10. Result: Experiments with network depth, convolutions, and latent_dim to minimize loss. Stack: Python, PyTorch Code: Jupyter Notebook

Objective: Analyze user funnel progression for a subscription-based product, evaluate A/B experiments with statistical significance testing. Stack: Python, pandas, statsmodels (z-test) Code: Jupyter Notebook

Objective: To analyze venture capital investments, startup funding activity, acquisitions, and investor behavior using structured SQL queries on a relational database. Data: Startup Investments dataset including company, fund, funding_round, investment, acquisition, people, and education tables. Stack: SQL, PostgreSQL Code: SQL Scripts

Objective: To build and train a machine learning model for forecasting the number of taxi orders in the next hour. Data is here. Code: Jupyter Notebook

Objective: Develop a method for encoding client data to make it difficult to restore private information. Data is here. Code: Jupyter Notebook

Objective: To analyze whether a client's marital status and the number of children they have impact their ability to repay a loan on time. Data is here. Code: Part 1 and Part 2

Objective: Conduct an Exploratory Data Analysis and calculate the profit. Data is here. Code: Jupyter Notebook

Objective: To build a machine learning model for predicting car prices to aid in setting accurate prices within a lead generation application. Data is here. Code: Jupyter Notebook

Objective: To develop a machine learning model to predict the coefficient of gold recovery from ore, addressing challenges from declining ore grades and ensuring economically viable gold mining. Data: Full data is here, and test data is here. Code: Jupyter Notebook

Objective: To analyze income tax data from different states in the US to uncover trends and insights related to tax collection and personal consumption expenditures. Data is here. Code: Jupyter Notebook

Objective: To identify which tariff brings more revenue. Data is here (please do not delete from the disk). Code: Jupyter Notebook

Objective: Build a model with the highest possible accuracy. The goal is to achieve at least a 0.75 accuracy score. Test the accuracy on the test dataset. Code: Jupyter Notebook. Sorry, I cannot find the initial dataset for this.

Objective: To analyze historical data on game sales, user and expert ratings, genres, and platforms to identify key patterns that contribute to the success of video games. Data is here. Code: Jupyter Notebook

1. First, we have a person with an idea. 
2. Money beats the idea. 
3. Politics beats money. 
4. But – surprise! Technology beats politics. 

(Jan Stenbeck)