Middle Inżynier interfejsów inteligencji maszynowej

Building production-grade LLM applications, advanced RAG architectures, and autonomous AI agents. • Co-designed and scaled a production-grade internal RAG assistant (serving 500+ daily active users, cutting employee research time by 40% and processing 15K+ multi-format documents) supporting heterogeneous data parsing, hierarchical chunking, and metadata filtering to optimize semantic context delivery. • Implemented a high-performance retrieval pipeline (dense vector search + BM25 + RRF + Cross-Encoder) that eliminated hallucinations on critical queries and boosted search accuracy from 62% to 89%, directly improving user adoption and retention. • Collaborated closely with product managers, data analysts, and backend engineering teams to align LLM capabilities with business requirements, translating complex user needs into robust technical specifications. • Contributed to the design of an offline evaluation framework with ~800 curated golden queries, utilizing Ragas to decouple retrieval evaluation (Context Precision/Recall) from generation evaluation (Faithfulness, Answer Relevancy) for automated CI/CD regression testing. • Implemented end-to-end LLM observability via Langfuse and OpenTelemetry, tracking P95 latency, token consumption, and failure modes, which drove prompt caching and context pruning strategies reducing inference costs by 30%. • Fine-tuned Llama-3.1-8B-Instruct via Hugging Face and QLoRA on 1.5K instruction examples using custom chat templates, boosting structured-response adherence from 78% to 91% and eliminating manual post-processing/formatting by support agents. • Developed core components of an agentic workflow integrating LLM tool/function calling with internal REST APIs and microservices, orchestrating multi-step execution loops (data retrieval, runtime validation, and action execution) with automatic error recovery and state management.