Kompletni AI stack pro cesky realitni trh. Od scrapingu 184 portalu pres 7-engine ocenovani po automatickou detekci prilezitosti.
184 portalu pres Data Plane v2.0 (warm pools: 3x Crawl4AI SMART, 1x Playwright HEAVY, httpx FAST). Sreality Direct API — 22K zaznamu za 7 minut. 20-thread image scraper (36K obrazku/8min).
Deduplikace, normalizace adres, RUIAN geocoding, image scraping (OG:image), parsing atributu. 65 sloupcu v MASTER_LIVE.db.
Super-Ensemble v2.1 — 7 nezavislych enginu: prodejni data, AI_V8 (XGB+LGB+GB+CatBoost), ocenovac (KNN 145K, 31ms), mapa_levon, smart, vynosovy, nakladovy. + Market Intel korekce.
Porovnani inzeratni vs. odhadnute ceny. 8,016 nemovitosti >30% pod trhem. Automaticky ranking, top prilezitosti na Znalecky Panel (3 AI znalci, ~4min/req).
Autonomni AI agent. MindSearch pattern — rozlozi dotaz na sub-queries, paralelne prohleda DB + web.
User: "Najdi 3+kk byty v Praze pod 5M s parkovanim"
|
v
MindSearch Planner (Qwen3:32b GPU)
| rozlozi na 2-4 sub-queries
|
+-- DB Search x3 --> MASTER_LIVE.db (217K)
+-- Web Search x3 --> Scraper (26 portalu)
+-- Merge + Dedupe --> LLM Combined Analysis
|
v
Vysledek: 12 bytu, serazeno dle opportunity_level
deck.gl vizualizace, glow efekty, sidebar s kartami. Mapa okoli s Turf.js radius.
Real-time odhad pres Super-Ensemble. Typ, mesto, plocha = 7 engine prumer.
Prime prohledani Sreality + dalsich portalu. Cerstva data mimo DB.
| Engine | Port | Metoda | Presnost |
|---|---|---|---|
| Prodejni data | :5091 | Historicke transakce, CSU | Baseline |
| AI V8 | :5093 | XGBoost + LightGBM + GradientBoost + CatBoost | Vysoka |
| Ocenovac | :5300 | GPS/KNN, 145K zaznamu, 31ms | Vysoka |
| Mapa Levon | :6200 | Proprietarni spatial model | Stredni |
| Smart | :5098 | Feature-weighted regression | Stredni |
| Vynosovy | :6002 | Kapitalizace najemneho | Investicni |
| Nakladovy | :6003 | Reprodukcni + opotrebeni | Pojistny |
3 AI znalci (Qwen3:32b) + RAG + MCP + konsenzus. Pro top prilezitosti. ~4min/request.
:5112Trzni trendy, sezonni korekce, lokalni faktory. Doplnkovy scoring.
:5400Denne 07:15. 4 listy, 50 sloupcu, color-coded prilezitosti.
deck.gl fullscreen s glow efekty. Radius, sidebar karty, MapTiler.
Real-time metriky pipeline. Pocty zaznamu, uspesnost scrapingu.
"Najdi byty v Praze pod 4M" = AI rozlozi, prohleda, vyhodnoti.
Nove mega prilezitosti (>30% pod trhem) okamzite notifikovany.
Automaticke generovani profesnich reportu s grafy a mapami.