AI Scientist

AI Scientist — OmniMind Laboratories

San Francisco, CA | January 2024 – Present

• Developed and fine-tuned a 7B-parameter large language model (LLM) using LoRA, improving domain-specific translation accuracy by 18% and reducing manual annotation costs by 32%.
• Optimized TensorRT inference engines for real-time video summarization pipelines, reducing API response latency by 35% and increasing throughput to 2,500 requests per second.
• Designed and ran a distributed multi-GPU training infrastructure using PyTorch and Ray, cutting model training time from 14 days to 4.5 days for a dataset of 50M images.
• Implemented automated drift monitoring and outlier detection workflows in production, decreasing post-deployment performance degradation incidents by 44% over a 9-month period.

Machine Learning Engineer — Neurala Robotics

San Jose, CA | June 2020 – December 2023

• Built a convolutional neural network (CNN) pipeline for automated defect detection, achieving a 98.4% precision rate and reducing quality assurance inspection times by 26%.
• Scaled AWS SageMaker training endpoints to run parallel hyperparameter tuning jobs, cutting model development cycles by 3 weeks and lowering cloud resources spend by $45,000.
• Collaborated with the software team to rewrite performance-critical preprocessing pipelines in C++, reducing data loading latency by 52% and speeding up epoch time by 18%.
• Architected a vector database indexing system using Milvus for high-dimensional embeddings search, reducing average search latency to 12ms for a corpus of 10M records.

Deep learning, Machine learning, Large language models (LLMs), Computer vision, Natural language processing (NLP), Distributed training, Neural network optimization, Model deployment, MLOps, PyTorch, TensorFlow, Hugging Face, Python, C++, AWS SageMaker, Ray, Milvus, Docker, Kubernetes, Git, Triton Inference Server

Distributed LLM Training Framework

Role: Lead AI Scientist

Tools: PyTorch, Ray, AWS EC2, Hugging Face

Architected a distributed fine-tuning pipeline for 13B-parameter models, reducing cluster communication overhead by 28% and completing training 4 days ahead of schedule.

Real-Time Embedded Computer Vision

Role: Machine Learning Engineer

Tools: TensorFlow Lite, Docker, C++, NVIDIA Jetson

Designed a lightweight edge inference model for object detection, achieving 45 frames per second on hardware with a 92% reduction in memory usage.

• Google Cloud Professional Machine Learning Engineer (2023)
• AWS Certified Machine Learning – Specialty (2021)
• NVIDIA Deep Learning Institute Certificate (2022)

• Languages: English (Native), Russian (Conversational)
• Volunteer Work: Technical mentor for regional AI and robotics youth workshops (2021-present)
• Availability: 3 weeks notice