An automated valuation model agents were willing to quote from

The challenge

Listing data arrived from eleven sources in eleven shapes, was cleaned by hand in spreadsheets, and reached the model weeks late. Agents had quietly stopped quoting the automated estimate because it was wrong often enough to be embarrassing, which made the whole feature decorative.

What we did

1. 01Replaced the manual clean-up with orchestrated ingestion per source, each with its own schema contract and quarantine for rows that fail it
2. 02Modelled the warehouse in dbt so every figure has tested lineage back to the source row that produced it
3. 03Split training and inference into separate pipelines with versioned datasets, so a model can be reproduced months later
4. 04Added freshness, drift and prediction-interval monitoring, and taught the product to withhold an estimate rather than show a bad one
5. 05Published confidence alongside every valuation so agents could see when to trust it