At Booking.com, we operate one of the largest Kafka infrastructures in Europe — spanning on-prem clusters, AWS MSK, and Confluent Cloud, supporting 1000+ engineering teams. With thousands of brokers and topics, Kafka admins face constant load: triaging user issues, enforcing best practices, managing ACLs, debugging consumer lag, and dealing with monitoring and deployment across various environments. During an internal hackathon, we built an AI-driven Kafka assistant to address these operational pains. Unlike generic copilots, our assistant was purpose-built for Kafka admins. It connects to Booking’s internal ecosystem (via MCP server), integrating with Slack and tools like Kafka Controller, Prometheus/Grafana, and internal deployment and ACL systems. Instead of context switching between dozens of tools, Kafka admins can now query one assistant to: – Check topic configs or replication states – Understand consumer lag anomalies – Suggest ACL changes or validate them – Link MR changes to potential Kafka risks – Get live cluster health summaries by combining metrics from different sources Under the hood, we use RAG (Retrieval-Augmented Generation), indexing Kafka playbooks, alert runbooks, and internal operational documentation. We iterated on indexing strategies to fine-tune the assistant’s performance — learning how RAG behaves with real-world Kafka content. We tested multiple LLMs and prompts but kept the assistant grounded: its goal isn’t to generate code or hallucinate answers, but to support Kafka workflows in a reproducible, auditable, and team-visible way. Finally, we’ll reflect on the compliance and auditability challenges we faced — such as access control— critical for production readiness. Whether you're operating Kafka at scale or exploring how AI can support your data platform teams, this session offers hands-on lessons from the field — not theoretical dreams.