Piotr Szczurowski

• • Architected and deployed a production LLM-powered conversational AI assistant handling 50,000+ daily interactions using LangGraph and OpenAI GPT-4, reducing support ticket volume by 38%

• • Built end-to-end RAG pipeline integrating Pinecone vector DB with enterprise knowledge base, improving answer accuracy by 45% over baseline retrieval

• • Designed and fine-tuned domain-specific NLP models using LoRA on LLaMA-2 for financial document classification with 94% F1 score

• • Implemented MLOps infrastructure using MLflow and Weights & Biases, reducing model deployment cycle from 3 weeks to 4 days

• • Led migration of batch inference pipelines to real-time streaming architecture using Apache Kafka and AWS SageMaker endpoints

• • Developed multi-intent NLP classification system using BERT fine-tuning, deployed on Kubernetes with sub-100ms p99 latency

• • Built conversational AI chatbot using Rasa framework integrated with Dialogflow fallback, achieving 87% user satisfaction score

• • Designed A/B testing framework for model evaluation across 12 concurrent experiments, enabling data- driven model selection

• • Implemented data pipelines processing 5TB+ monthly using Apache Spark on AWS EMR for feature generation and model training

• • Delivered predictive analytics models using scikit-learn and XGBoost for client churn prediction (AUC 0.91 ) and revenue forecasting

• • Built EDA and visualisation dashboards using Python and Plotly for stakeholder reporting across 8 enterprise clients

• • Prototyped early conversational AI experiments using rule-based NLP pipelines, forming the foundation for later LLM integration work