Ivan Poleschyuk

My blog about machine learning, computer vision, and other programming stuff.

I have analyzed the metrics produced by vertebral column scanning machine of novel architecture (3D scanning and motion capture based) with R. I had anonymous data of repetitive visits of more than 500 patients and used ggplot2 to visualize correlations between metrics for repetitive visits of the same patient versus data of unrelated patients. My report had outlined which metrics are most useful for patient diagnostics and which are too noisy or are unlikely to be linked to any medical issue.

The goal of the project was to build a model, which will produce an additional input to the trading system (along with technical indicators) built by a private client for a day trading on a MICEX futures market. I have used R (dplyr, ggplot2, neuralnet, betategarch, xts, ttr) to work on this task. The main challenge, as always with financial time series, was the dynamic nature of the process modeled. GARCH was used to estimate the optimal time window to train the model for predictions. It also turned out that simpler models, like logistic regression, converge better on such data.

This amusing project was about estimating a weight of each of several small animals in the cage (exact number unknown) using data feed from one IoT scales in each cage. I used Python (pandas, scipy, numpy, matplotlib) to visualize the data and work along with the client on an optimal clustering-based algorithm to determine the exact weight of each animal. The main challenge was the fact that there was no way of telling if animal climbed on the scales entirely.

In this project, the goal was to store and analyze the data from hundreds of CCTV traffic cameras across the city to detect rush-hour bottlenecks, count number of vehicles, detect speed limit violations. I've built PySpark scripts to collect these statistics over an Apache Spark cluster. The data was stored in CSV format across multiple Hadoop nodes, several hundred gigabytes in total.

This is the music recommendation service I've built as a hobby project exclusively on open source technologies.

59th (top 6%) - Walmart Recruiting: Trip Type Classification. This competition involved complex market basket analysis. I have used R to preprocess (dplyr) and explore the data (ggplot2), then run a forecast with random forest and xgboost, and fed preprocessed dataset to neural net model built with Python to ensemble predictions.

In this project, I have worked as one of computer vision developers to build a Microsoft HoloLens augmented reality application to assist the surgeon in a knee replacement surgery. The goal was to eliminate some of the hospital's costs associated with buying expensive guiding devices and replace them with holograms. Main technologies: C++, OpenCV, UWP, and HoloLens.