The task (Real-time Bidding) was to predict the bet to the online auction of SSP. The problem was in the speed of the model inference and predicting bets when do not reduce the winner cases. Hard load system. The Model in production and Increased profit by 10%

Fraud detection model with a low time lags and use the behavior of Publishers. CatBoost is used. Active learning to get new labels of Fraud on each iteration. PoC

Traffic optimization is the on-line model of a relevant traffic classifier (‘Good fo DSP). Model in production. Improve traffic by 5% CatBoost is used.

The task was to classify gases using the electric signals & wavelet transform. Upgrade of the working model. AWS solution

The task was to classify different statuses of the broadcasting. The problem was in a few data set. I solved it by active learning way labeling of new data. AWS solution, Sagemarket for production. Transfer learning. The accuracy on the test set was 0.9.

The task was to detect the different behaviors on the stairs, needs to insurance tasks. The problem was in a few data set. I solved it by separating the task into two sub-tasks, first to extract features from images and second to use these features to predict the behavior. The extraction was solved by CNN which was trained by the images to see the important features and after then I feed RNN this feature to predict the behavior. This way I've could use my laptop for this task, only 16 Gb RAM. The accuracy on the test video set was 0.9

ask was to get the most possible text recommendation for different people. The initial data were different states of activity monitor and environment data. Solved by factorization machine method.

First I’ve done Outlier detection. Next, I was working on a feature extraction problem and feature engineer task. Then I used clustering to build the different models to different clusters. And last I've tried to build the boosting model and added something like a factorization machine to the model. In my work I use Numpy lib and write the optimization code from start to end, in the last cases I use the TensorFlow and PyTorch. In general, I've made the improvement to an accuracy of 2%. Model in production