Senior AI & ML Engineer

I own and develop production LLM-agent capabilities, including the Hypothesis MCP Server built with FastMCP. I implemented the tool framework and 11 of 32 tools covering contradiction detection, document retrieval, agent memory, and knowledge-graph writing. I also delivered Kubernetes autoscaling, Langfuse tracing with per-call LLM cost tracking, and blocking security scans in CI.

I built agent memory from scratch using hybrid BM25, vector, and graph retrieval with reciprocal-rank fusion, episodic LLM compression, and cross-session inheritance. I developed MCP tool-calling agents with ReAct loops and OAuth2 authentication, invented contextual RAG embeddings, and fine-tuned embedding models. My work also includes computer vision, multi-object tracking, sports analytics, LLM fine-tuning, automotive R&D, security remediation, and MLOps delivery on Kubernetes and Google Cloud.

I built production real-time bidding and advertising ML systems for Sift mobile. I developed per-campaign CatBoost click and conversion models retrained daily from S3 logs, integrating fraud blacklists and automated deployment to a DSP prediction API.

I designed install-prediction, win-probability, bid-price, and bid-optimization models using CatBoost, custom focal and asymmetric losses, SageMaker pipelines, Docker, Helm CronJobs, Redshift, and ClickHouse. The delivered systems improved traffic filtering and winning-bid efficiency at very large scale.

I developed an AWS-based computer vision system from scratch for a multi-camera vehicle-analysis use case. The solution integrated three camera streams, spatial object detection, and vehicle data fusion.

I implemented calibrated coordinate mapping, cross-view object association, and vehicle type and color identification. The work required combining computer vision models with geometric calibration and multi-view tracking logic.

I built the AI core of AiBro, an application that automatically replied to users' Facebook Messenger chats. The system retrieved relevant content from each user's chat history using BERT embeddings, cosine similarity, and triplet-loss Siamese ranking.

I also developed standard-answer and trigger-phrase layers with 365,000 template variants. The BERT response ranker improved Recall10@3 from 0.69 to 0.84.

I developed a real-time broadcasting-status classification solution from a small training dataset. I used active-learning labeling and transfer learning to improve model quality under limited data conditions.

The production solution achieved 0.9 accuracy and was deployed on AWS SageMaker. The technical stack included TensorFlow, Keras, and OpenCV.

I delivered machine learning projects for external clients in sensor analytics and recommendation systems. One project classified gases from sensor signals using wavelet transforms and was upgraded for production on AWS SageMaker and AWS Lambda.

I also developed a recommender system using feature-interaction models, Vowpal Wabbit, and gradient boosting. The work involved practical model development and cloud deployment for US-based clients.

I developed production models for horse-race winner prediction. The approach combined outlier detection, per-cluster models, boosting, and factorization-machine terms.

The resulting system improved prediction accuracy by 2% and was deployed in production. My work focused on feature engineering, model selection, and improving predictive performance on structured data.

I worked on data science and software development projects across text analytics, classification, topic modeling, and anomaly detection. I built text classifiers spanning 150 classes on 8 GB of text and web-user classifiers covering 550 classes.

I also worked with LDA and BigARTM topic modeling and performed anomaly search related to the 2014 Ukrainian election. This period established my foundation in applied machine learning and large-scale text processing.

PhD in Aerodynamics and Numerical Methods.

Master of Science in Applied Mathematics and Physics.