Main developer of a chat-bot assistant with Retrieval Augmented Generation (RAG) over corporate documents:• Implemented RAG from scratch, using Azure Functions, Cognitive Search, Blob Storage, WebApp, BotService, BotBuilder python SDK and LangChain, accelerating accurately finding the corporate information for the customer's employees by several times• Significantly increased the quality of a baseline RAG Q&A by implementing various Prompt Engineering techniques, such as: few-shot, Chain-of-Thoughts, conversation history management, etc.• Improved quality (NDCG score) of the Information Retrieval by experimenting with various advanced retrieval methods such as HyDE, Context Reordering, Long Context Reordering, Parent Document Retrieval• Drastically reduced hallucinations by proper prompting and implementing Attribution Score calculation for Q&A validation• Developed PII recognition and masking pipeline that allowed customer to securely store metadata of conversation with the Bot• Developed a content extraction service, which allowed to extract images, tables and text from .PDF, .DOCX and images using Azure Functions and Document Intelligence

Developing a chatbot assistant with Retrieval Augmented Generation (RAG) over domain-specific knowledge base with Large Language Models (LLMs):• Significantly reduced the backend latency on the production by developing a backend architecture with Redis for Q&A cache and multiple async invocations of several LLMs; • Helped the customer save production costs and latencies by replacing OpenAI GPT-3.5-turbo with open source classification model for Intent classification without losing classification quality;• Drastically improved the quality of Q&A by implementing proper prompting techniques, such as kNN for few-shots, Chain-of-Though, well-structured prompt, and dynamic prompt usage based on the user's intent;• As a Data Science team leader, I closely collaborated with Frontend, QA, Backend, DevOps and Product teams to coordinate the development of the final solution;• Coordinated work processes of 2 other Middle Data Scientists, specifying tasks and scopes of their work, helping and supporting them;

I took a role as the main developer of a NER model over legal documents:• We achieved beating the customer's main competitors' and AWS Comprehension quality on NER task• I setup EC2 instances with PAWLS labeling tool, configured and established deliverance of the labeling results to S3 and from it to ML pipelines• Written most of the classes and DS/ML code for data preprocessing, training of HuggingFace model, validation of the results with search-and-filtering task specifics, postprocessing of the model's output• Delivered ~10 NER models with SageMaker to the production• Constantly analyzed model's errors and mistakes to improve the results and described its behavior to the business • Constantly communicated with the business side to clarify the metrics, needs, and expectations from the model and the whole ML pipeline• Covered DS/ML code with unittests using pytest libraryConducted 20+ experiments during 6 months including: • data augmentation with customers and open-source data• finetuning of 10+ different transformers architectures• back-translation dataset augmentation• heuristics based on the business logic over predictions• entities manipulation: relabelling, combining, removal, etc.

Hierarchical Text Classification:• Proposed and implemented a baseline solution for a multilabel classification model based on the token extraction from the NLP data, this baseline was successful and allowed us to deploy an MVP very soon;• I was responsible for deploying and establishing MLFlow for model performance review and validation, which allowed our team to move fast in our experiments;• Implemented a Hierarchical Multi Label Classification approach based on recent (at that moment) papers, improving the baseline model quality;• Incrementally updated the model with new classes maintaining and improving the overall quality with hierarchical models, built with HuggingFace python API • Discussed, planned, and established an active learning approach with the Data Analytics team, managed the process, and reviewed the results of the labeling done by assessors;• Was deeply involved in the communication with the customer regarding sprints' planning, explanation of the model's behavior and its outputs, showing the demo and business processes clarification • Took responsibilities as a team leader during his absence or high load: tasks deadlines estimation, decomposition, delivering, demos, team management• Wrote multiple tests with pytest for different functions, methods of classes

People Analitycs project:• Proposed new workflows for 3 streams on the project - StarScore, Assessment Score and Attrition Score (all of them are regression tasks), suggested and wrote the project's new repository class structure• Performed EDA of multiple data sources (.csv), which allowed the team to improve the feature engineering pipelines and unblocked new possibilities there;• Implemented feature importance analysis (with Shap values), helped the business understand the behavior of the models and identify major factors to adjust business processes;• Successfully, proposed, tested and deployed a new hypothesis for replacing the current target with a new one generated and sampled with an approach based on time-series analysis;• Wrote and optimized numerous .sql scripts (PostgreSQL) for ETL from multiple sources of structured and semi-structured data, as well as for feature engineering, decreasing the production execution latency by several times;• Refactored legacy source code (Python) for an attrition score prediction, made a clearful and extensible class structure, and integrated into the existing solution;• Closely communicated with the business side to specify the requirements, discuss possible features to insert into the existing pipeline, show and analyze feature importances, and gain feedback on the model's output and behavior.

• Performed Exploratory Data Analysis of the dataset of semi-structured, highly sparsed data of various products• Wrote a pipeline for filtering and converting text data into various vector representations, such as BytePairEmbeddings (BPE), and BPE weighted by Tf-Idf scores.• Experimented on the clustering of the product variants with HDBSCAN, Hierarchical clustering and a weighted fully connected graph to identify groups of similar products to be shown as recommendations on the PDPs• Performed outliers analysis and wrote an algorithm for assigning them to clusters based on simple probabilistic heuristics• Closely collaborated with UI team to establish the design of the widget on a PDP where the results of the clustering will be shown to bring value to the customer• Wrote validation pipeline, clustering pipeline and unittests on the backend

NER model for parsing candidates' skills from CVs to accelerate and improve staffing processes in the company: • Discussed the customer’s requirements and vision of the final product• Wrote an ETL pipeline with parsing and processing of candidates' CVs from various sources • Performed EDA of various data formats: PDF, DOCX, TXT• Experimented on the NER model for parsing skills from CVs and established the final model based on the validation results