Senior Инженер по интерфейсам машинного интеллекта

• Support Ticket Intent Classification & Routing: Deployed an 8-class intent classifier (BGE-M3 + CatBoost, ~3.2K tickets) with Regex NER for order_id extraction, achieving 94% Accuracy, 0.91 Macro-F1, and ~95 ms p95 latency while cutting triage time to 1–2 seconds. • Production Lead Scoring System: Built an end-to-end lead scoring pipeline combining cargo deion embeddings (LaBSE-ru-turbo) with tabular/CRM features via CatBoost (~10K deals). Automated instant Telegram alerts for high-priority leads, achieving 88% Accuracy and 0.92 ROC-AUC. • Uplift Modeling for Promo Campaigns: Co-developed a customer selection model using a T-Learner architecture (CatBoost) to predict order probability with/without discounts. Deployed via FastAPI to CRM fields, achieving a +5 p.p. conversion boost (20% vs 15%) and reducing discount expenditure by 27%. • Internal RAG Assistant & Agents: Scaled production RAG serving 500+ DAU (15K+ docs, reducing research time by 40%) using dense/sparse hybrid search (Qdrant + BM25 + RRF + Cross-Encoder) that boosted accuracy from 62% to 89%. • Agentic Workflows & SFT: Developed multi-step agentic loops via MCP/FastMCP and fine-tuned Llama-3.1-8B-Instruct using QLoRA on domain datasets, increasing structured-response adherence from 78% to 91%. • MLOps & Evaluation: Built offline evaluation pipelines using Ragas (800 golden queries) and integrated end-to-end observability via ClearML and Langfuse, supporting continuous retraining and pruning optimizations that reduced inference costs by 30%.