PropMetrics

Overview

A real estate data platform that scrapes the 48 neighbourhoods of Buenos Aires every night and turns the listings into decisions: price per m² by neighbourhood, yield and payback, an algorithmic appraiser with a confidence interval, an opportunity detector, short-term vs. traditional rent, flipping margins, and prices deflated by UVA, CER or ICL to see what actually moved. It ships its own dependency-free dashboard and answers the same questions over WhatsApp through Respondi.

The Challenge

Argentina's real estate market has no transparent data: listings are scattered across portals, prices are asked prices, and inflation makes any historical series unreadable. On top of that, the portals actively block datacenter traffic — the first scraper worked locally and returned nothing at all from the server.

The Approach

Rewrote the Spring Boot backend as an async FastAPI service with Alembic migrations and an arq worker for the nightly jobs. The blocking turned out to be measurable rather than absolute: one portal only filters IPv4, another rate-limits after six requests, so the scraper paces itself, backs off, and cools a portal down instead of giving up on the whole run. Every metric ships its sample size and returns null when the data is not there yet — that is what let the analytics land before the historical series existed.

The Outcome

Live on its own domain with the nightly pipeline running: 48 neighbourhoods, cross-portal deduplication, geocoding, appraisal models retrained daily, and a market digest that arrives on WhatsApp every morning. The appraiser sits at 21.5% median error on sale and 16.9% on rent, with its p10–p90 interval covering 81.5% of held-out listings.

Key Highlights