Bohdan – Deep Learning, Data Science, NLP

Leading presale activities - from requirements analysis and scoping to team composition, implementation and hand-off of POCs/MVPs to the client mostly focusing on projects involving LLMs from simple RAG to CUA.

• Implemented CUA/vLLM workflow for data processing purposes, handling errors of the existing parsing RPA;
• Optimized token expenses by 20% using prompt compression tools;
• Implemented code generation focusing LLM applications for data exploration and test automation using Assistants API;
• Developed LLM applications for fuzzy-matching and data enrichment pipelines processing invoices.

Aerial imagery analysis project focusing on real estate property analysis involving a cascade of computer vision models providing more than 140 insights about the object of interest to insurance companies.

• Migrated a pipeline of U-Net models to AWS SageMaker multi-model endpoints, reducing cloud infrastructure expenses by 10x;
• Retrained YOLOv8 models resulting in 5x inference speed up as well as zoom level invariance;
• Implemented SAHI approach to accommodate inference on large images (up to tens of thousands of pixels);
• Led and coordinated release activities with the DevOps team.

B2B ad tech project, mostly focused on real-time bidding models and NLP-based ad campaign ranking system with millions of daily queries and managing/monitoring available ad inventory.

• Developed ad campaign/website ranking system using the latest NLP models for millions of websites and thousands of campaigns daily, optimized storage costs, and employed heuristics and algorithmic optimization;
• Developed model monitoring tools, dashboards, and back-testing instruments;
• Conducted A/B testing of RTB models and different features
• Reduced bot traffic by 1.5%;
• Took part in GPT-3/4/J, Chat-GPT POC development.

Internal securities data provider with conversational AI capable of text2code generation (including complex aggregate requests), basic dialogue capabilities, embedding-based search engine, and a recommender system.

• Developed a hybrid approach to search engines/content-based recommendation systems, measured model effectiveness, beating Elastic Search nDCG score by 70%;
• Developed conversational AI models capable of translating user utterances into SQL/NoSQL requests;
• Created synthetic datasets from scratch, used data augmentation techniques;
• Took part in speech recognition R&D;
• Fine-tuned BERT-based NLP models to specific domains.

An AI start-up is trying to monetize ML in every conceivable way. From predicting sports outcomes and tracking players from video feeds to reinforcement learning trading agents.

• Created a sports outcome prediction model and paired it with a profitable bet allocation strategy;
• Took part in the development of tracking through detection model (background subtraction + Kalman filter) for football player analysis, optimized reading stream, improved FPS throughput from 30 to 45 fps;
• Developed several Q-learning-based trading agents, beating the annual return benchmark by 1.5%.

Open source python clustering library.

• Was in charge of a visualization task of genetic algorithm clustering during his own research;
• Used PyClustering as an open-source library; however, during his experiments, he noticed incorrect crossover mask generation and certain issues with the constructor parameters;
• Fixed the bug with mask generation;
• Slightly refactored the constructor so it was able to have the required parameters passed correctly;
• The changes were introduced manually in PR #474.

IP telephony/B2C ads focused CRM with Google analytics/social media integration and basic conversion outcome predictive models.

• Was responsible for REST API development;
• Was in charge of oAuth2.0 integration;
• Prototyped and integrated a regression model for conversion modeling using Accord.Net;
• Created clustering based on client website activity;
• Integrated speech-to-text translation of recorded phone calls into the CRM using Google Speech API.