AI Property Valuation in Buenos Aires, Where Nobody Publishes the Sale Price

In Buenos Aires, every apartment has two prices: the one on the portal, and the one it actually sells for. Nobody publishes the second one.

So sellers improvise. The asking price is often sentimental: a round number the owner won't go below, because it feels right.

Then the price drops a couple of thousand dollars every few days until the listing is "burnt": stale, and "hard to sell".

So I built Tasalo. Describe your apartment and get its value, a rental analysis and a PDF report. A weekend idea that grew to 30 commits, a Mercado Pago checkout (Argentina's default way to pay), login by emailed link, coupons and 46 neighbourhood price pages. It's live, in Spanish, at tasalo.tulocal.ar; quotes from it are my translation.

Where does the data come from when the portals block you?

From the portals' own search pages, fetched slowly from inside a real browser, after two funnier methods failed.

One of the big listing portals sits behind a web application firewall, a bot blocker. Plain requests worked for a few dozen pages, then came back as HTTP 202 with an empty body, a polite "prove you're human": the firewall's challenge, not an answer. My first scraper read that as "no listings" and cheerfully recorded zeros. A neighbourhood with zero apartments. In Buenos Aires.

Attempt two: copy the firewall's token cookie from a real browser and replay it. It worked exactly once. The firewall fingerprints the connection itself, so a valid cookie from the wrong kind of client is still a stranger.

What works: the fetch runs inside a real Chromium, requests two to three seconds apart, a fresh browser context every 25 or so, any challenged context binned on the spot. Search cards already carry price, size and rooms, so no single listing ever gets opened. Fewer requests, fewer enemies. Three workers were fine. Four unpaced ones got the whole run blocked after about 150 requests.

Bonus comedy:

• The count went backwards. Two scrapers shared one checkpoint file, each overwriting it from its own stale copy. The "dead" one I'd relaunched was only napping in a retry backoff.
• The government spells creatively. The city's price file says "Paterl" for La Paternal and "Boca" for La Boca. Without aliases, three neighbourhoods quietly lost their history.

What isn't honest yet?

Six things, listed, because a valuation tool that hides its weak spots is just a confident stranger.

• The history line is mislabelled. My pages call the city's price series closing prices. The statistics office's own reports say it's built from a big portal's listings, so it's asking prices, like mine. Correction owed.
• The "real value" is a guess with decimals. Tasalo shows a market value 20% below asking. That 20% has no source. The only closed-deals index I found puts the haggle near 5%.
• The five-year projection is compound interest in a suit. It's the neighbourhood's yearly trend, compounded five times.
• Two datasets, not one. The evaluator still runs on an older 15-neighbourhood price table; the pages run on fresh medians for 46.
• The homepage brags. It shows a big round "properties analysed" number that my own runbook says the data doesn't back.
• La Boca outranks Recoleta. The page admits why: its listings skew to renovated lofts on the tourist circuit. A median is honest about its sample, not about the city.

And the testimonials: three glowing, star-rated quotes on the homepage, all placeholders. They're gone from the live page. A testimonial is a lie until a customer says it. Real usage: a handful of evaluations.

Here's the whole engine minus the PDF, fed with the neighbourhood pages' medians, the two discounts side by side. Play with it: