Sources and methodology

Methodology

• Neighbourhoods. 905 neighbourhoods (məhəllə) across Baku, Sumqayıt and Gəncə: the 67 with a boundary in OSM (microdistricts, settlements) are used as they are; the rest of the area is filled with Voronoi cells built from OSM settlement points and metro stations and clipped to district boundaries, so boundaries are approximate (not aligned to streets). Indicators are aggregated from an internal H3 grid (3,700 cells): counts are summed, distances are area-weighted averages and prices are medians; distances to the nearest metro (checked against the 27 operating stations), stop, school, kindergarten and park; names come from OSM settlements. The price is the median if a neighbourhood has ≥5 listings, otherwise the district average.
• Decision engine. The AI structures the query and explains the result; code filters and ranks matching areas and computes the confidence score (coverage 40%, query completeness 25%, gap between 1st and 2nd 20%, match with constraints 15%).
• Areas. OSM administrative boundaries: 75 districts/cities and Baku's 12 districts. Sea areas are clipped by the Caspian coastline (islands kept). Population and density: SSC, 01.01.2026.
• Services. OSM features (pharmacies, groceries, schools, kindergartens, hospitals, clinics, banks, cafés, sport, parks etc.) are counted if they fall inside the polygon and divided per 10,000 residents.
• Transport. Metro and railway stations (unique by name), bus stops, bus/minibus routes mapped in OSM; for Baku districts, the straight-line distance from the centre to the nearest metro.
• Roads. Main roads, streets and rural tracks; surface from the OSM “surface” tag: asphalt/concrete, dirt/gravel or unknown. On the map, “paved %” is calculated among roads with a known surface.
• Nature. Land cover is sampled from OSM polygons on a 30×30–36×36 point grid in each area (approximate). Elevation: Open-Meteo, 5×5 grid; terrain: plain (mean <200 m, std <50), mountain (mean >1000 m or std >250), otherwise foothills. Climate: ERA5 (2025); PM2.5: CAMS model.
• Indices (0–100). Each indicator is converted to a percentile among areas at the same level, and the index is their average: comfort, transport, education, healthcare, nature, road infrastructure, affordability (inverse of price).
• Fit. The AI extracts weights from the persona sentence; score = Σ(weight × index) / Σ weight. If the home exceeds the budget, the score is multiplied by 0.3 and marked “over budget”.
• True cost (10 years). Home price + 120 months × (monthly travel cost + hours travelling × value of an hour). Distance = straight line × 1.4; speeds: metro 30, bus 15, car 25 (Baku) / 60 km/h. Approximate model.

How the AI works

• The AI runs on Google Gemini (flash). The model sees only the area's JSON indicators and may use only the numbers in it; every answer is validated against a JSON schema (zod).
• Area profiles are generated in advance and cached; persona, Q&A, comparison and legal check are called live and the result is cached.
• When the AI is unavailable (rate limit, network), the site falls back to a rule-based answer and says so in the text.
• In the legal check, known risk phrases (“çıxarışsız”, without a title extract; “kupçasız”, without a deed; “etibarnamə ilə”, sold under power of attorney, etc.) are also detected by a deterministic rule, so the AI cannot miss them. A law article is cited only if its text is in our database.

Limitations

• Prices: bina.az listing medians in 17 areas (2024, a user-supplied dataset; we do not scrape and store aggregates only), 7 published figures, the rest demo. Most districts have no price.
• OSM is incomplete in the regions: few features means “not mapped”, not “not there”. The coverage level is shown for every area.
• Route counts reflect only lines mapped in OSM; in the regions this figure is very low.
• PM2.5 is a model estimate, not a measurement; climate is for the area's centre point.
• In Karabakh and East Zangezur, population and infrastructure are changing fast, so the data may lag behind.
• There is no congestion or real travel time; this would come only through an official Yandex/Waze partnership (no scraping).