Problem
Energy market data (NEM dispatch intervals, generation by fuel type, curtailment events) is structured, time-sensitive, and technically dense. Let non-technical users query it in natural language without the AI fabricating numbers or citing stale data.
Approach
Built the query interface as an AI-to-API bridge, not an AI-to-training-data system. Claude translates natural language into structured API calls against live ClickHouse data, then interprets the returned result into plain language. The model never generates a data point — it only interprets one.
- AI translates intent → API call → interprets result.
- No model-generated statistics — all figures from live query.
- Generative UI: charts rendered from query results, not described.
- Ambiguous queries ask for clarification rather than guessing.
Result
Public-facing query interface where a policy researcher can ask "how much of SA's power came from wind last Tuesday?" and get an accurate answer with a chart — not an approximation from training data.