LeadVillas: finding villa buyers on Bali and Phuket in 45 languages
People looking to buy a villa ask in their own language, in chats, comments and forums all over the world. LeadVillas reads 35 sources in 45 languages, keeps only real buyers and investors, and delivers each request newest first, translated, with a first message drafted in the author's language.
- Languages
- 45, plus any other via the model
- Sources
- 35 and growing
- Stack
- Python, SQLite, Claude, Telegram
- Runs
- Around the clock on a server
What it does. A dictionary-based classifier in 45 languages flags intent to buy or invest, price questions under competitors’ listings, budgets in any currency and contacts, and filters out sellers and agents. Claude then checks each candidate in any language, writes a short summary and drafts the first message in the author’s language. Place names are recognised in every alphabet and case form.
One base, no duplicates. The same request posted in ten chats becomes one lead marked as actively looking. Leads are ranked by score times freshness, with an 18-hour half-life, so the newest serious buyers arrive first.
How it is checked. A built-in critic scores the dictionaries, the classifier on an independent corpus of 450 posts in 45 languages written by another author, a separate held-out set, source coverage per language and live source health. Buttons under each lead (“take it”, “not a lead”) feed quality back per source.
The rule it keeps. It never messages anyone automatically: it hands over the contact, the link and the draft, and a person replies.
My part. I designed and built it: sources, the multilingual classifier, the Claude check, ranking, the Telegram bot and the critic.