Csaba Szegedi 🇭🇺

Large Language Models are moving fast into regulated, sensitive workloads — but every prompt, response, and trace you log for debugging is also a potential compliance liability. How do you get the observability your engineers need without turning your logs into a GDPR/HIPAA incident waiting to happen? In this talk, we'll walk through a real-world AWS architecture built to solve exactly this trade-off: a data-masking and tokenization pipeline that sits between your application and Amazon Bedrock, using AWS Batch for processing, Amazon S3 and CloudWatch for storage and monitoring, and TLS/encryption best practices throughout — all designed so sensitive data is reversibly masked for privacy while traces stay useful enough to actually debug production issues. You'll leave with a practical pattern for privacy-by-design observability on AWS: what to mask, what to preserve, how to choose between AWS-native logging features and custom pipelines, and how to configure different observability "risk levels" for dev vs. production. Whether you're building GenAI features for healthcare, finance, or any data-sensitive domain, this session gives you a concrete blueprint for shipping LLM features your compliance team won't lose sleep over.