From fine-tuning GPT-OSS for multilingual reasoning to building human-in-the-loop systems and better ways to evaluate RAG.