"Show me all pod crashes in the payments namespace in the last 10 minutes." Your junior engineer typed that in plain English and got an instant answer – no PromQL, no Lucene syntax, and no tribal knowledge required. In this lightning talk, I'll walk through the production architecture of an NLP-to-SQL engine built on top of ClickHouse + OpenTelemetry, handling petabyte-scale telemetry data in real workloads. The core challenge: translate natural language reliably into optimised ClickHouse SQL across a dynamic OTel schema, validate the output, and return sub-second results.

I'll cover:
- Prompt engineering for SQL generation (LLaMA via Groq) with runtime schema injection
- Query validation and fallback strategies to guard against hallucinated SQL
- Dynamic field mapping from OTel semantic conventions to ClickHouse columns
- Real edge cases: ambiguous queries, multi-service correlation, time range inference
- Performance patterns for querying billions of log rows without breaking latency SLOs