VienenRaíces

Houses, flats and land, for sale and for rent. The difference is the search box: you type what you need in your own words — area, budget, bedrooms — and a model turns it into structured filters (Claude with structured outputs, an offline heuristic fallback in both languages, and a cache keyed by the query so repeated searches cost nothing). Results are ranked by a score that blends semantic similarity from pgvector embeddings with a fit score for the things you asked for, price closeness and text completeness; if nothing matches, a second relaxed pass widens the net and says so.

Underneath is one Supabase project — Postgres 17 with pgvector and PostGIS, 19 tables, 40-plus row-level-security policies. The app only ever holds the publishable key: authorisation lives entirely in the database, and anything privileged goes through narrow security-definer RPCs. A Python worker trains a gradient-boosting price model on log price with quantile bands, writes an estimate and a verdict — below, in line with, or above the market — onto every active listing, and computes a demand index per district from searches, views and leads.

Around that: listing import from a URL, feeds from agency CRMs, saved searches with alerts, appointment booking with anti-overlap constraints, a report-style listing page with price analysis and zone insights, and a split results view with a synchronised map. Live at vienenraices.com.