CV

Remote, USA

• Founded and architected a managed agent platform for reliable, sandboxed execution of containerized agents at 10,000+ concurrent containers, with multi-launcher support across Cloud Batch, GKE, and Modal.
• Architected and operated a distributed evaluation-as-a-service platform used by external frontier-model customers to run reproducible evaluations against model checkpoints, with CI/CD, experiment tracking, automated reporting, and real-time performance dashboards.
• Unified 12+ external and internal reasoning, coding, knowledge, and agentic benchmark suites — including HLE, AIME, Tau-Bench, MCPUniverse, WebArena, SWE-Bench, Terminal-Bench, LiveCodeBench, SciCodeBench, and BigCodeBench — under a single evaluation harness for consistent cross-model comparison.
• Founded and served as primary architect of a multimodal RL Gym and sandboxed execution platform, enabling managed agents to complete tasks across unified GUI, browser, and CLI environments; reduced recurring environment startup time by ~50%.
• Built reusable evaluation tooling, configurable scoring libraries, and LLM-rubric workflows — including sandboxed agent runs, judge-agent grading, pass@k majority voting, weighted scoring, and per-criterion reports.
• Co-authored a software-engineering benchmark for agentic use cases, leading task definition, dataset sourcing and curation, coding-data collection, evaluator implementation, and QA.

New York, USA

• Built and productionized an LLM-powered semantic product-search application on Amazon Bedrock, SageMaker, PostgreSQL, and pgvector, achieving 25ms p95 retrieval latency.
• Improved search relevance by 15% to 0.96 nDCG@10 through embedding fine-tuning and results reranking.
• Led a generative-AI initiative from concept through integration, contributing 25-50% of long-term company revenue.

New York, USA

• Implemented Two-Tower Embedding Model with PyTorch achieving nDCG@10 of 0.9 on the Amazon ESCI dataset.
• Conducted empirical studies showing ~15% accuracy and f1-macro improvement for LLMs across diverse datasets.
• Improved SOTA metrics for BERT for hate speech detection in terms of plausibility by ~3%, faithfulness by ~4.5%.

Greensboro, USA

• Applied transformers for NLP and NLU, enriching historical data and improving users' retention prediction by ~3.5%.
• Built DataBricks pipeline using PySpark for automatic monthly data cleaning, processing and model fine-tuning.
• Conducted survival analysis on time series achieving a c-index of 0.92 and ~96% accuracy with Deep Learning RNN.

Kharkiv, Ukraine

• Developed distributed computing solution processing 10x larger datasets, enhancing system throughput by 40%.
• Designed Redis-based cache for the CRM system with 20 million records, reducing average request latency by ~25%.
• Authored the system with ~$2.5M USD live market cap, 100+ active validators, and ~1000 transactions per second.
• Supervised resource utilization analysis, optimizing model inference processes resulting in ~45% latency reduction.

New York, NY — GPA 4.0/4.0

Thesis: Context Attribution To Attain Generalizability in Non-fine-tuned Transformer: A Framework for Concept Extraction in Cross-Dataset Evaluation Settings (PDF), advised by Dr. Ruhul Amin

Kharkiv, Ukraine — High Honors, GPA 97.74/100

Thesis: Smart Drone System for Road Quality Assurance (frontend) — a CNN-based (TensorFlow) drone and IoT system for real-time road quality monitoring and alerts