Alexey Gavrilov Email and Phone Number

- 2 Anti-Fraud classification tasks on tabular data: Made data analysis, univariate and multivariate analysis, tuning hyperparameters and feature selection, built final models- Mentoring Intern for clustering on tabular data: Intern made data analysis, feature preprocessing, research different clustering methods, described fraud and normal clusters portrait.The work was carried out in close contact with an expert from the customers.Mentee successfully solved this task.

Some of the tasks: - Covenant Date Monitoring Parser: Covenants monitored by the bank have a monitoring start date. The goal is to findthis date from covenant text. For this task, it is necessary to correctly split the covenants into types and extract the valuesspecific to these types.Used Sentence RuBert with F1 0.97 for classification task, trained T5 model with accuracy 0.98 for extraction task.Model deployed as Restful-Api microservice- Reviews Sentiment Classifier: All reviews with must be classify as neutral, problems and proposals. The main difficultythat problem text often overlaps with proposals and that this classifier must be universal for most of Sber services.This task was solved with macro F1 0.8- Types of Realty Corrector: Contractors can make mistakes in realty types when entering data. The goal is to correctthem and find outliers at data.Made feature engineering, solution based on catboost has macro F1 0.98.Model is deployed as a service that is activated by the data mart on a schedule

Full cycle of ML development: communication with the customer, task definition, data mining, datapreparation, model creation, testing, deployment- Built prototype of IFRS reporting OCR&NER, which reduces manual work for two times- Factor analysis, research for improving early warning system in credit monitoring. Business impact of this model is over 5 billions per year- Some tasks on Data Analysis for building an employee profile

Сделал прототип классификатора изображений под АндроидИспользовал алгоритмы классического комьютерного зрения, генеративные подходы для создания синтетических обучающих данныхИспользовал методы активного обучения для кратного увеличения реальных данных