Anton Potapov

AI Engineer with 6+ years of experience building production NLP, LLM, and RAG systems for customer experience analytics and financial intelligence. Specializes in multi-agent orchestration, retrieval and model fine-tuning, and multilingual NLP. Hands-on experience spans data curation, evaluation, evidence-grounded generation, and inference optimization, delivering measurable gains in retrieval quality, analytical efficiency, and recommendation performance

• Agile
• AI-assisted development
• Analytical
• Souci du détail
• Tests automatisés
• CI/CD
• Collaboration
• Communication Skills
• Continuous Improvement
• Critical Thinking
• Analyse de données
• data management
• Data Validation
• Visualisation de données
• Decision Making
• Detail-oriented
• Docker
• Documentation
• GitHub
• JavaScript
• Kubernetes
• Leadership
• Apprentissage automatique
• Mentorship
• Microsoft Azure
• Microsoft Excel
• Microsoft Office
• Next.js
• Sens de l’organisation
• Pipeline
• PostgreSQL
• Power BI
• Résolution de problèmes
• Process
• Process Improvement
• Production
• Gestion de projet
• prompt engineering
• Python
• Assurance qualité
• React
• Collecte des exigences
• API REST
• Root Cause Analysis
• Software
• SQL
• Tableau
• Travail d’équipe
• Gestion du temps
• Workflow Automation

1. AI Engineer

   Qualtrics

   août 2022 to Aujourd’hui

   ● Designed a LangGraph / Llama 4 Maverick multi-agent root cause analysis (RCA) system combining Qdrant hybrid retrieval with Snowflake journey analytics. Implemented ReAct planning, analytics, hypothesis generation and validation, confidence gating, and structured reports with source attribution. Reduced analytics preparation from several hours to ~20 minutes and improved problematic journey detection accuracy by 28%. ● Built an enterprise RAG platform using Qdrant, BM25 + dense retrieval, metadata filtering, cross-encoder reranking, and DeepSeek V3. Fine-tuned e5-large with contrastive learning and bge-reranker for ranking on ~6K query-document pairs with hard negatives. The combined retrieval improvements raised Recall@10 from 0.71 to 0.86 and reduced analytics preparation time by 43%. ● Developed a LangGraph / DeepSeek R1 multi-agent copilot for NPS/CSAT feedback, with query decomposition, Qdrant retrieval, deterministic complaint aggregation, and evidence-grounded summaries and recommendations. Implemented validated tool calls, retries, and fallback handling, reducing preparation from several hours to ~40 minutes. ● Fine-tuned Llama 3 70B with 4-bit QLoRA and supervised instruction tuning on curated customer feedback examples for executive summaries and recommendations. Improved response quality by 24% in human evaluation and factual consistency by 27% versus the prompting/RAG baseline. ● Built intent, churn-signal, and sentiment classification across 14 languages using DeBERTa v3 and XLM-RoBERTa, achieving macro F1 ~0.93. Applied language-based model routing, focal loss, hard example mining, and probability calibration to support proactive retention workflows. ● Built human-led evaluations supported by LLM-as-judge, groundedness checks, and MLflow tracking for models, prompts, agent traces, and quality drift. Prepared Dockerized inference services and API contracts for MLOps deployment; implemented PII masking and schema-validated, restricted tool access.

2. AI Engineer

   AlphaSense

   avr. 2020 to août 2022

   ● Built financial entity and relation extraction with spaCy, BERT, dependency parsing rules, and entity/ticker normalization. Extracted companies, tickers, M&A transactions, forecasts, and risk factors, achieving entity-level F1 ~0.91. ● Fine-tuned FinBERT for earnings-call and financial-news sentiment analysis, aggregating company-level signals for downstream trading and risk models; improved trading signal precision by 18%. ● Built financial-news topic modeling with transformer embeddings, BERTopic, UMAP, and HDBSCAN, using analyst validation and topic-drift monitoring to identify emerging market themes and anomalous news activity. ● Developed personalized content ranking with XGBoost and LightGBM using behavioral, content, and temporal features. Addressed implicit-feedback and position bias; evaluated ranking with NDCG and Precision@k and validated a 21% recommendation CTR uplift through A/B testing.

1. Torrens University Australia

   Master of Business Information Systems

   Adelaide, Australia

   2020 to 2022

• anglaisProfessionnel