Mikhail Gurevich

A Python-based agrotech project which provides agriculture companies management with information on the quality of soil tillage. I was responsible for the development of an analytical service that:
- Gathered information from a number of sensors installed on the agriculture equipment (plows, cultivators, tractors): GNSS data, accelerometers, gyroscopes, etc.
- Worked with this data as a time series. This step involves filtering out noise from the data and normalize the data across the time axis (different sensors have different data frequencies).
- Analyzed the data using heuristics alongside decision trees and linear classifiers
- Prepared reports for management using Seaborn and Plotly
- Used PostgreSQL as storage for raw, normalized, and processed data
- Used pandas, NumPy, and GeoPy to process the data

A Python-based application that generates images of food based on the recipe and ingredients list. I was responsible for choosing neural network architecture (GANs) for our model and took part in the experiments during the model training process.

I also proposed and implemented a novel technique based on a few papers from https://arxiv.org/. We used multimodal learning for the generated image quality improvement. Particularly, we built representations in one feature space for texts and images and then trained the model using triplet loss to classify (text, image) pairs.

Later on, this classifier was used as an additional term in the generator loss.

A PyTorch-based end-to-end model that finds all the car plates on the photo and recognizes text on them. The dataset contained 25,000 real-life photos of cars in different cities. Each photo could have contained several cars.

I made a pipeline which consisted of the following parts:

1) Mask R-CNN (https://arxiv.org/abs/1703.06870), a model that detects all car plates presented in the photo.
2) Preprocessing of car plate images using OpenCV library, adjusting blur and the contrast of the images in order to normalize them before passing them to OCR.
3) Char-RNN (https://arxiv.org/abs/1706.01069) a model specifically designed for OCR.
4) An additional step was made to increase the quality of OCR based on the knowledge of the car plate's text structure. I implemented and trained a language model based on the Beam Search.

This model achieved a Levenshtein Mean of 1.05 on the test dataset. As the company disclosed afterward, the test dataset contained car plates from different regions, so test data had a different distribution from the train data.