2020 - 2023
2022 - 2023
KEY FUNCTIONS: ➤ Full ownership, development and adoption of DL / ML NLP approaches within an Agile product team ➤ Collaboration with management and business stakeholders to identify directions for product growth ➤ Designing, conducting, and analyzing production testing of launched AI features ➤ Management of a team of 2+ ML Engineers / DSs KEY ACHIEVEMENTS: ➤ Reduced operators' Average Handle Time by 2% by developing and deploying from scratch Sber’s first MVP intent-based classifiers for text and voice channels ➤ Achieved T2M of 1 day for production model releases by automating an end-to-end process for model training / validation and deployment ➤ Launched from scratch a data labeling process for both channels ➤ Played a pivotal role in persuading business stakeholders to approve roll-out of the first text AI Bot despite mixed production test results, by showing in-depth analytical proof that operators would still benefit ➤ Received a special honor from Sber’s CEO for exceptional success in mentoring junior specialist
2021 - 2022
KEY FUNCTIONS: ➤ Development and adoption of ML / DL response models for non-bank products’ sales personalization ➤ Large-scale customer data analytics to identify behavioral patterns leading to a purchase ➤ Prototyping large-scale features’ datamarts to fasten the process of their production implementation KEY ACHIEVEMENTS: ➤ Enhanced response to Sber’s subscription marketing banners by 2% by developing a set of GBDT models to personalize sales, launched as a Hadoop-backed Batch Service to score >100m clients in a single run ➤ Acknowledged as the most pro-business approach for non-banking products’ sales personalization and selected by Sber’s Head of Retail for a deep-dive into modern ML approaches for response modeling
2020 - 2021
KEY FUNCTIONS: ➤ End-to-end development, support, and production implementation of DL / ML approaches to enhance Sber's intent-based ChatBot's resolution rate ➤ Direct collaboration with business and data annotators teams to ensure the proper quality of labeled data and tailor the Bot to current business needs ➤ Mentoring of interns and junior specialists KEY ACHIEVEMENTS: ➤ Boosted performance of client's intent classification model (BiLSTM + Attention) by ∼10% by ensuring rigorous control for labeled data and customizing feasible Keras RNN / CNN model architectures ➤ Decreased T2M for production releases from 2 months to 2 weeks by unifying modeling and deployment pipelines and establishing a regular collaboration process with business stakeholders and data annotators ➤ Reduced the length of clients’ paths to the correct answers on 2 steps for ~10% of requests by developing scalable rule-based slot-filling pipelines for clients' intents of any complexity
2018 - 2020
2018 - 2020
KEY FUNCTIONS: ➤ End-to-end ML / DL models development, production implementation, business effect evaluation, performance monitoring and support ➤ Analysis and quick prototyping of existing market solutions / technologies / open data sources to assess their applicability for the Bank ➤ Prototyping large-scale features’ datamarts (in Oracle) to fasten the process of their production implementation KEY ACHIEVEMENTS: ➤ Deployed into production > 5 ML / DL projects from different business domains ➤ Generated > 50% of Data Science Department NPV ➤ Promoted from Junior to Senior role in 1 year and 2 months PROJECTS EXAMPLES: ➤ Clients’ payment patterns embeddings: Custom RNN-based aggregation of client’s payment history from Credit Reference Bureaus / client’s transnational data to predict the probability of her default. Implemented as a Flask REST-API online service. Enhanced ROC-AUC of the Bank’s ScoreCard by∼8%. One of the most powerful features of the ScoreCard ➤ Legal Debt Collection ScoreCard: GBDT ScoreCard to automate the selection process of clients in deep delinquency to be sued. Implemented as an Airflow Batch Service. Boosted yearly debt collected per sued client by ∼50%. The most profitable project in a department ➤ Interactive Voice Response System customization: A set of GBDT models for different IVR branches to determine reasons for clients' calls to the bank to re-play them at higher positions in IVR. Implemented as a Sanic REST-API online Service. Reduced time spent by clients to get information of interest by ∼10% ➤ Agents fraud detection model: RNN-based aggregation of agents’ applications to detect “suspicious” ones. Implemented as an Airflow Batch Service. Improved the number of detected fraudulent agents by ∼20%
2014 - 2014
2014 - 2014
2013 - 2013
2013 - 2013
2012 - 2013
2012 - 2013