Nikita Barinov

Developing generative AI for advertising and LLM & multimodal systems for VK Video, one of Russia’s largest video platforms with millions of daily users.

Generative AI for Advertising | July 2026 – Present

• Built a production pipeline that transforms static banners into video ads through automated scriptwriting, image generation and video synthesis.
• Replaced template-based generation, increasing clicks by 33% and conversions by 19% while reducing CPC by 21% and CPA by 11%, with only 5% higher ad spend.
• Optimized MiniMax H3 (TI2V) inference with vLLM-Omni, achieving ~4× faster generation in benchmarks through attention backend optimization and denoising-step tuning, and reducing per-GPU memory usage by 28% with tensor parallelism.
• Implemented asynchronous generation workflows with Temporal, integrating model APIs, artifact storage and analytics.

LLM & Multimodal AI for VK Video | October 2025 – June 2026

• Fine-tuned LLM embedding models for retrieval, improving quality on internal relevance benchmarks.
• Designed a reproducible experimentation pipeline and conducted 100+ training experiments, expanding relevance dataset coverage and diversity.
• Developed multimodal content models, improving recall by 14% at a fixed precision of 0.90.
• Distilled OpenAI o3 into a compact relevance LLM, achieving consistency recall of 0.93 and precision of 0.95 on internal evaluations.
• Deployed production models with Triton Inference Server, using TensorRT and ONNX to optimize inference.