Senior AI & ML Engineer

I own and develop production LLM and MCP capabilities, including the Hypothesis MCP Server built with FastMCP. I implemented the tool framework and 11 tools covering contradiction detection, document retrieval, memory, and knowledge-graph writing, while supporting Kubernetes autoscaling, Langfuse tracing, per-call LLM cost monitoring, and CI security scanning.

I built hybrid agent memory from scratch using BM25, vector, and graph retrieval with reciprocal-rank fusion, episodic compression, and cross-session inheritance. I also contributed to multi-agent orchestration, contextual RAG, LLM contradiction detection, embedding fine-tuning, sports computer vision, tracking, and LLM fine-tuning projects, and was part of the SoccerNet GSR 2024 winning team.

I developed production machine-learning models for mobile advertising, including per-campaign CatBoost click and conversion models retrained daily from S3 log windows. The bid-filtering solution improved precision to 3.7 times the base click rate and achieved ROC-AUC of 0.80.

I designed install-prediction, win-probability, bid-price, fraud-detection, and bid-optimization models using CatBoost and custom asymmetric and focal losses. My work supported high-volume DSP traffic, SageMaker pipelines, Docker and Helm-based deployment, Redshift data processing, and an exchange deployment filtering 19% of traffic while increasing winning bids per request.

I built an AWS-based computer-vision system from scratch using three cameras, spatial object detection, and vehicle-data fusion. The system included calibrated coordinate mapping, cross-view object association, and vehicle type and color identification.

I worked on integrating vision outputs across camera views into a unified spatial understanding of detected vehicles. The engagement focused on production-oriented architecture and robust multi-source data fusion.

I built the AI core of an application that automatically replied in 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 implemented response-ranking, standard-answer, and trigger-phrase layers supporting 365,000 template variants. The BERT response ranker improved Recall10@3 from 0.69 to 0.84.

I developed a real-time classifier for broadcasting statuses using a small dataset. I applied active-learning labeling and transfer learning to improve the effectiveness of limited training data.

The solution achieved 0.9 accuracy and was deployed into production on AWS SageMaker. The implementation used TensorFlow, Keras, and OpenCV.

I delivered applied data-science solutions for clients in the United States. Projects included gas classification from sensor signals using wavelets and a recommender system using feature-interaction models and gradient boosting.

The gas-classification solution was deployed on AWS SageMaker and AWS Lambda. I used Vowpal Wabbit and machine-learning approaches suited to production inference and client-specific data.

I developed horse-race winner prediction models using outlier detection, per-cluster modeling, boosting, and factorization-machine terms. The work focused on improving predictive quality in a production setting.

The resulting approach increased accuracy by approximately 2%. I applied feature engineering, clustering, and ensemble learning techniques to racing data.

I worked on a range of machine-learning and software-development projects, including large-scale text classification, topic modeling, web-user classification, and anomaly detection. My text-classification work covered 150 classes and approximately 8 GB of text.

I also developed web-user classification systems with 550 classes and performed anomaly analysis related to the 2014 Ukrainian election. The projects involved NLP, LDA, BigARTM, and scalable data-analysis methods.

Completed a PhD in Aerodynamics and Numerical Methods.

Developed advanced foundations in numerical modeling, applied mathematics, optimization, and scientific research.

Completed a Master of Science degree in Applied Mathematics and Physics.

Built a strong foundation in mathematics, physics, probability, numerical methods, and computational problem solving.