Semen Poliakov

Brief list of my achievements in Intercom for the first year: - Accelerated backtesting LLMs: general speed up 3x for all models, speed-up for specific use-cases: ~50x - Spearheaded LLM fine-tuning (LoRA), with ~8-15% improvements for groundedness and helpfulness metrics - Developed several versions of the Voice AI assistant core - Radical improvements for configurations flexibility of the RAG system, pipelines and knowledge base

I went from Middle to Lead ML Engineer in 3 years, and finished my journey in VK.com as a Team Lead of Content Analysis group, responsible for most of Deep Learning models that we used in core Recommendation systems: - Led development and deployment of dozens of ML models (RecSys, NLP, CV) in production, including a flexible similar-post search system, thematic recommendation feeds, and multimodal content uniqueness detection, boosting user engagement across VK. - Drove measurable business impact, increasing site-wide comments by 5%+, enhancing user engagement metrics (views, likes >5%, timespent). Developed and launched from scratch a set of thematic feeds reach to over a million users. - Spearheaded infrastructure and system improvements, from building an MVP Feature Store and updating ML pipelines to developing debugging tools, high-throughput ML backend (3000 qps), and modernizing classification systems. - Mentored and led high-performing teams, winning 1st place in an internal hackathon among 50 strong teams, guiding interns to publish articles and defend theses, and pioneering ML advances with cutting-edge embedding models (fastText, ELMo, BERT) and published papers about Active learning and Normalizing flows.

Over as short period of time, developed: • Time series analysis and anomaly detection systems in Python • Service in Java to manage analytics module