Toronto Open Data Assistant

Residents ask a plain question. The system finds the right dataset, reads it, and answers with the numbers and the source next to each other.

The hard part was never the model. It was ranking, chunk boundaries, and knowing when to say the data does not cover this.