Mikhail Nefedov Email & Phone Number

Amsterdam, North Holland, Netherlands

Moscow, Russian Federation

Designed a course on Natural Language Processing. Taught 3 iterations of the course to more than 70 students (developed jupyter-notebooks, gave lectures and seminars, reviewed home assignments, conducted the exams). The course was highly rated by the students (one of the best in several categories in 2019 and 2020).Collaborated on developing an introductory course on Machine Learning. Taught 5 (out of 8) topics of the course. Reviewed home assignments and gave personal feedback. The course was highly rated by the students.

Moscow, Russian Federation

Worked on several different projects:AI-assistant for a call-center. Implemented a classification model which predicts (in real time) a topic of the call using its transcription. The model helped the operators to log the details of the conversation more efficiently. Voice-assistant. Implemented high-precision grammar-based and matching-based models for extracting a diverse set of named entities (such as dates, intervals, periods, tasks, contacts).Chat-bot platform. Developed a compact deep learning model for text vectorization using knowledge distillation which was used for intent classification and semantic deduplication. The model allowed to automatically answer 15%-20% of user questions to support service. Also analyzed the logs to find weak spots of existing classifiers and applied clustering methods to discover frequently asked questions.Agency for Health and Consumer Rights. Developed several models for the Russian state regulator to automate fraud detection in distribution of products of animal origin across the whole country. The complexity of the models varied from complex regexps and large postgresql queries to gradient boosting classifiers. Opportunity scoring for CRM system. Currently working on a model that aims to predict the probability of closing a deal with a contractor. The challenge here is that there are few historic data available and that the model needs to be robust and interpretable.

Moscow, Russian Federation

Designed metrics to measure search quality and user satisfaction. The metrics were instrumental in discovering problematic topics that needed legal experts’ attention. The metrics were also used to identify clients likely to churn.Developed vector-based search engine for specialized search for judgements of Russian courts of law. The engine used word embeddings and approximate nearest neighbor search. The resulting product allowed to quickly search through millions of judgements using either short or very long queries (it was possible to use whole judgment text as a query to find court positions on similar situations)