Property market intelligence platform
Turning a government property-transaction firehose into a queryable intelligence layer that powers verifiable market reports.
Problem
Property-transaction records lived in PDFs and slow government portals. Every market claim a content team wanted to make was either unsourced or took days to verify by hand, so reports were thin, late, and impossible to audit.
Constraints
Source data is messy and revised retroactively. Zero tolerance for fabricated numbers: every figure has to trace back to source. Solo build, running on commodity hardware, no managed data warehouse budget.
System
An ETL pipeline normalizes registry CSVs and a commercial market-data feed into an embedded DuckDB warehouse, cross-referenced against BigQuery. A validation layer recomputes every metric from source data before any figure is allowed to reach content, so nothing reaches a report that cannot be traced.
Outcome
The platform makes 1.6M+ transactions queryable in seconds and guarantees that every figure in a generated report traces back to source data, so nothing reaches a client that cannot be defended.
Stack
Python and FastAPI for the service layer, DuckDB as the embedded warehouse, BigQuery for cross-referencing, and the Claude API for drafting on top of validated data.