Buy Houses Japan: find property in Japan using AI

Buy Houses Japan is a live bilingual property search app built and operated by Urbano DX. A buyer types one sentence, the app turns it into editable structured filters, and results land in a card list next to a map with price pins. Coverage spans Kanto, Kansai, and Kyushu, with EN and 日本語 interfaces, USD or JPY prices, and sqft or m2 units.

Open Buy Houses Japan

The filter panel after a plain-sentence search: four AI-derived filters (Osaka, house, max price, minimum 3 bedrooms) shown as editable fields next to the manual filter form

Results for that query: 429 listings found with 4 active filters, shown as a card list beside an interactive map with price pins

A listing detail page: normalized fields from the Japanese source listing (layout, ownership, year built, nearest station, zoning, areas in m2) with a floorplan gallery, relevance score, and market read

What it is

A buyer writes something like "quiet family home near a train station in Osaka with a garden under 30 million yen". The app parses that into structured filters, shows them as editable fields, and runs the search. That exact query returns 429 matching listings across 18 pages, with prefecture, property type, maximum price, and minimum bedrooms filled in and open to correction.

Plain-sentence search parsed into editable filters
429 results on the sample query, paginated over 18 pages
Card list beside a map with price pins
Bilingual EN and 日本語, USD or JPY, sqft or m2

What was built

Listings come from major Japanese portals through scheduled ingestion pipelines, then get normalized into consistent fields: property type, layout such as 3LDK, ownership, year built, nearest station and transport access kept in Japanese, zoning, and building and land area in m2. On top of that, each listing carries derived signals: a relevance score, a market read against comparable prices, and condition flags.

Front end: React 18 and Vite, built with Bun, maps via MapLibre GL
Back end: FastAPI on MongoDB, with zvec for semantic search
Ingestion: Prefect pipelines scraping AtHome and Nifty
Firebase for auth, Stripe for payments, LLM calls through a shared gateway

What it demonstrates

The pattern travels well beyond real estate. Messy user intent becomes structured criteria the user can see and override, noisy source data becomes normalized records, and the model output stays advisory rather than hidden. Urbano DX built, shipped, and operates the whole surface, which is the same discipline applied to paid PoCs and MVP sprints.

Intent to structured filters, always editable
Messy Japanese source data normalized into usable fields
Derived signals shown as signals, not verdicts
A live product to inspect, not a slide deck