Artem Budishchev

A DRL model trained to optimize a client's KPIs.

I led the project completely end-to-end from its inception—researching and prototyping, deploying to production, and monitoring the performance. I supervised the software development team, which was supporting the project from a technical point of view. The model itself was implemented using Python and TensorFlow and deployed using Kubernetes on AWS.

A deep contextual bandit trained to optimize a client-provided KPI.

I led the project end-to-end from ideation, researching and prototyping, deployment to production, and monitoring the performance. A customer's state feature representation was transformed by an autoencoder and subsequently clustered. The clusters were used as contexts in the bandit. The model was implemented using Python and TensorFlow and deployed to Kubernetes on AWS.

A fraud detection model that was developed for the Spanish market.

I collected, parsed, and cleaned the data from various sources (databases, flat files, cloud storage, and so on). We built a CatBoost-supervised learning model, which showed a healthy performance in the backtest. The model was trained to directly optimize a KPI, which led to a significant increase in monetary gain.

I created a message (emails and SMS) classification model using a fine-tuned BERT model from Google, which significantly outperformed the baseline model (logistic regression and bag of words). Feel free to reach out for more details.

I developed a sophisticated attribution modeling system that consists of four different approaches based on Shapley value, logistic regression, Markov chain model, and XGBoost model. The final attribution was calculated using the ensemble's prediction median.

I performed analytics in order to cross-sell and upsell customers across various lending products (POS, installment loans, and so on).

The project produced lists of customers likely to purchase the offered product or upgrade the existing one.