u
u_2a2be8a02b26

Mykola M

@u_2a2be8a02b26

LLM Research Engineer

Czech Republic
Ukrainian
About me
I am an LLM Research Engineer with practical experience in developing, training, and optimizing large language models (decoder-only transformers) from scratch and for product needs. I specialize in self-supervised learning, distributed training (DDP/FSDP), fine-tuning methods (LoRA, QLoRA), and integration of LLMs into agent systems and RAG applications. I have a strong background in PyTorch, data engineering, and building production-ready AI solutions.... Read more

Skills

u
u_2a2be8a02b26
Mykola M
Offline • 
Average response time: 2 hours

See my services

AI Integrations
I will develop advanced ai ml solutions or agi prototypes
From$550 / projectMore details

Portfolio

Work experience

ML Engineer

Free

Mar 2022 - Present • 4 yrs 2 mos

- Developed and trained decoder-only LLM from scratch for internal AI products. - Conducted experiments with transformer architectures, attention schemes, and tokenization. - Implemented fine-tuning of models via LoRA / QLoRA with significant reduction in GPU memory consumption. - Optimized distributed training in a multi-GPU environment using DDP and FSDP. - Implemented mixed precision training (BF16), reducing training time. - Built ETL pipelines for processing large text corpora (SQL, Parquet, PyArrow). - Integrated LLM into RAG applications and agent systems (LangChain / LangGraph). - Configured monitoring of response quality and latency of models in production. - Presented research results and recommendations to product teams.

ML Engineer

Deep Lerning studio

Feb 2018 - Feb 2022 • 4 yrs

Project: Analysis and preparation of statistical reports - Set up data collection and analysis from NoSQL databases for the analytics department. - Created a dashboard using Plotly and Dash to monitor the company's KPIs. - Optimized database queries, reducing processing time by 30%. Project: Face and product recognition system - Developed a system based on PyTorch for real-time face and object recognition. - Implemented YOLOV5 for product detection on video, achieving 92% accuracy. - Created and trained a custom model for image segmentation using U-Net. Project: Video stream analysis and heat maps - Developed a module for analyzing the movement of video stream objects. - Used OpenCV and numpy to create heat maps based on video analytics. - Implemented a monitoring system in the retail network, which reduced queues by 25%.