In this talk, I present a real-world candidate pre-screening system, used as a case study to explore RAG architecture, data quality, and benchmarking challenges. The project includes real data engineering work designed to process resumes, prepare LLM-ready datasets, and rank professional profiles from CVs using RAG. The key aspect of this project is that the same solution is implemented twice, allowing a direct comparison of design, behavior, and results across two ecosystems: - Microsoft .NET (C#) using Semantic Kernel - Python using LangChain I walk through the end-to-end RAG pipeline: dataset preprocessing, vector indexing, skill normalization, and the generation of explainable responses suitable for hiring processes. This is not a theoretical talk. The session includes a live demo, real engineering decisions, and concrete problems encountered in practice such as data parity issues, inconsistent metadata, result quality problems, testing strategies, and performance trade-offs and how they were identified, benchmarked, and solved. The live demo focuses on comparing real queries, responses, and latency across both stacks, followed by benchmark results. RAG Repository: https://github.com/maurogioberti/llm-candidate-rag-benchmark-multilang What attendees will learn: - How RAG systems work and how to apply them to real Human Capital workflows - Practical differences between Semantic Kernel and LangChain in production-like scenarios - How to integrate .NET, Python, and LLM frameworks into a maintainable architecture - Best practices for data quality, testing, benchmarking, and result interpretation - Actionable patterns and anti-patterns for building and benchmarking RAG systems - Access to a fully open-source codebase for further exploration and learning

Session Speakers

Mauro Gioberti

.NET Software Engineer | Devs Mentor