Building LLM-powered applications, semantic search and RAG

* Develop proof of concept (POC) to validate more than 16,000 Short-Term Rental listings and enhance enforcement of STR regulations, supporting British Columbia's housing affordability goals* Utilize the LLMs (Gemini 1.5, GPT-4o, etc) and Deep Learning techniques to extract features and analyze host reviews, rental descriptions and images* Collaborate cross-functionally with engineering, product management and AI teams

1) Currently building Legal ChatGPT to help Canadian lawyers find legal cases based on sophisticated criteria faster2) Insurance Claims Automation:* Developed a PoC that simplifies insurance form completion (both audio and text versions) for a Canadian insurance company with a customer base of over 2 million clients. The PoC aimed to decrease customer wait time from 3h to 1s during peak hours and reduce call center costs by 87%.* Designed a backend architecture utilizing AWS API Gateway and AWS Lambda, integrated with the OpenAI GPT-4 API, to create a robust and scalable system for processing and managing form automation.* Infrastructure as code with AWS CDK, gTTS for text-to-speech, and OpenAI Whisper for speech recognition.* MLOps practices: CI/CD, staging, logging, pre-commit hooks, linter.3) Hotel information Q&A:* Built a hotel information Q&A bot using OpenAI GPT-3.5, LangChain, and Chroma RAG, based on touristreviews and chat messages. Simplified the process of checking if a hotel satisfies unusual requests.4) Bank Transaction Fraud Detection:* Solved Binary Classification for an imbalanced dataset of over 1 million transactions and identity information. Performed feature engineering using PCA and label encoding, trained LightGBM* Built inference infrastructure using API Gateway to access the service, 3 Lambdas to add incoming requests to SQS, process them, store them in a DynamoDB table and return predictions. Implemented using AWS CDK* Conducted payload testing5) Real Estate Rentals Analytics:* Directed the end-to-end ML lifecycle for a rental price prediction project, including data scraping for over 50 cities in Canada, feature engineering, model development, and maintenance* Managed XGBoost, Linear Regression models with MLFlow, used Postgres Database

* Implemented end-to-end pipeline (Python 3, PyTorch) for preview generation (solved as Video Summarization task) * Researched 4 SOTA approaches (CA-SUM, PGL-SUM, DSNet anchor based and DSNet anchor free), reimplemented and reevaluated original papers, modified architectures to achieve better results* The international team was supervised by Senior Computer Vision Research Engineer* Technologies: PyTorch, OpenCV, Docker, ONNX, GPU, AWS, moviepy, ffmpeg* Source code: https://github.com/anya-mb/summarization_models_inference

Solving Text Analysis, Demand Prediction, Scoring, Optimization tasks for international companies.Python, SQL, SAS.• Developed pilot ML models to predict contractors’ default rates for one of the leading steel manufacturers in Russia. It was expected to help our customer to personalize contract conditions, decrease accounts receivable by 15%, increase revenue by 12%.• Built pilot ML models to predict sales for petcare department of one of the leading international FMCG companies. Decreased SMAPE by 16% and increased department efficiency.• Developed pilot algorithms for automatic parsing of company’ agreement documents in Russian (extracting companies’ related information, representative names and job titles, prices). The project was expected to reduce the company’s Annual Salary Costs by 9% and to facilitate the creation of a user-friendly platform for search and comparison of agreements by selected criteria.