Inżynier ML

• Built an internal AI portal unifying 3 enterprise-knowledge capabilities-semantic search, context-aware Q&A, and solution discovery-across 28K+ documents and 150+ internal users, reducing information lookup time by 34.8%. • Developed a project-management analytics module integrating 7 data sources and tracking 8 team-health indicators, cutting manual analysis by 30 hours per month and surfacing delivery risks up to 3 days earlier. • Implemented a business-requirements agent that validated 12 completeness criteria, detected missing components, and enriched specifications using data from 4 internal systems, reducing review iterations by 32.4%. • Built an AI merge-request reviewer processing 80+ MRs per month and generating actionable findings, reducing review time by 26.7% and increasing pre-merge issue detection by 20%. • Standardized AI-assisted engineering across 3 SDLC stages-code generation, validation, and review-for 20+ engineers, improving delivery throughput by 18.7%. • Established an end-to-end content-moderation R&D pipeline spanning 4 stages from data preparation and model fine-tuning to evaluation and production deployment; achieved 0.88 F1-score on 12K labelled samples

• Developed LLM agents across 4 business functions-reporting, sales, operational analytics, and manager support-automating 8 recurring workflows and saving approximately 120 hours per month. • Built an executive AI agent that consolidated data from 7 internal sources, generated 6 recurring management reports, and flagged 5 deviation and risk categories, reducing the reporting cycle by 61.4%. • Developed a real-time call assistant processing 1,200+ calls per month; combined Whisper large-v3 / faster-whisper, Silero VAD, and pyannote.audio to extract 4 action-oriented outputs-key points, agreements, customer requests, and follow-ups-reducing post-call administration by 8 minutes per call. • Designed LangGraph/LangChain workflows across 7 orchestration capabilities-intent classification, tool calling, memory, routing, retrieval, fact-checking, and response validation-raising successful task completion to 86.2% and reducing unsupported answers by 32%

• Implemented a RAG pipeline for enterprise knowledge access across 100K+ documents and 1,000+ users using Qdrant, BM25 + dense retrieval, multilingual-e5, bge reranking, and Qwen-14B-Instruct; achieved 0.87 Recall@10 and reduced unsupported answers by 42.4%. • Fine-tuned the 13B-parameter Vikhr model with PEFT/LoRA for 3 use cases-classification, structured extraction, and style-aligned generation-improving average format compliance from 74% to 94%. • Developed and optimized a Russian zero-shot TTS model, improving synthesized-speech naturalness by 15.43% on Mean Opinion Score and reducing synthesis time by 14.72%. • Designed an LLM/RAG evaluation framework spanning 6 quality dimensions: relevance, factual consistency, faithfulness, numerical accuracy, citation quality, and hallucination robustness; evaluated 8 model and pipeline versions per release. • Built a RuBERT NER + rules system to extract technical parameters from 500+ PDF documents per month, achieving 0.91 F1-score and reducing technical-report preparation time by 65%. • Trained a customer-request classifier on 120K historical tickets using TF-IDF + Logistic Regression and later CatBoost + fastText, achieving 0.87 macro-F1 and automating 72% of departmental routing • Implemented an ASR and topic-modelling pipeline for 250+ hours per month of video-conference trans, migrating from LDA to BERTopic and improving topic coherence by 28% to surface recurring discussion pain points.