Dubai Property Investment Assistant

AI-powered investment scoring and Q&A for Dubai apartments, built on real Dubai Land Department open data: 47,677 registered apartment sales and Ejari rent contracts across 29 communities (January–September 2026). It computes rental yields, measures price momentum and supply pressure, scores communities, and answers questions with citations. A community without enough rent data gets "insufficient data", never an estimate.

Stack: Next.js 16 + FastAPI + DuckDB + ChromaDB + Groq (LLM). Bayut API optional for live listings.

Architecture

graph TB
    DLD[DLD open-data exports<br/>sales, Ejari rents, projects] --> Loader[dld_loader.py<br/>clean + map communities]
    Loader --> DuckDB[(DuckDB<br/>sales, rents, projects + views)]
    DuckDB --> Reports[build_reports.py<br/>one report per community]
    Reports --> ChromaDB[(ChromaDB<br/>market reports)]

    UI[Next.js 16 UI :3000] -->|REST| API[FastAPI :8000]
    API --> Router{Query Router}
    Router -->|Structured| DuckDB
    Router -->|Semantic| ChromaDB
    API -.->|Optional live listings| Bayut[Bayut API]

    DuckDB --> Scores[Investment Scoring]
    Scores --> LLM[LLM Generation<br/>Groq / OpenAI / OpenRouter]
    ChromaDB --> LLM
    LLM --> |Answer + Citations| UI

    subgraph "4-Factor Scoring"
        S1[Rental Yield 40]
        S2[Supply Pressure 25]
        S3[Price Momentum 20]
        S4[Liquidity 15]
    end
    S1 --> Scores
    S2 --> Scores
    S3 --> Scores
    S4 --> Scores

Quick Start

git clone https://github.com/ahmedbasolan/dubai-property-bot.git
cd dubai-property-bot

1. Export the data. DLD's open-data gateway requires a captcha per search, so the raw exports are a manual step (about 10 minutes). On dubailand.gov.ae/en/open-data/real-estate-data, download CSVs into data/raw/dld/, named transactions-*.csv, rents-*.csv and projects-*.csv:

Tab	Filters
Transactions	Sales · Residential · Unit · a date range in the current year
Rents	Registration Date · Residential · one month at a time (larger ranges can fail)
Project	Completion Date · current year

The gateway only serves the current year. Exact filters, column codes and what each export contains are in docs/plans/2026-10-23-real-dld-data.md.

2. Build and run.

make setup          # loader -> DuckDB -> community reports -> ChromaDB

cp .env.example .env
# Add at least one LLM key: GROQ_API_KEY (free tier), OPENAI_API_KEY or OPENROUTER_API_KEY
# Optional: BAYUT_API_KEY for live listings (RapidAPI)

make run-api        # http://localhost:8000  (docs at /docs)
make run-ui         # http://localhost:3000

Re-export and run make setup again to refresh. Adding more rent months fills in yields for more communities.

Docker (optional): docker compose up --build. The API container runs the same pipeline on start, so data/raw/dld/ must hold the exports.

Live demo: https://dubai-property-bot.vercel.app. One Vercel project serves the frontend and the API as two Services; see docs/deploy.md. The demo runs without a server LLM key, so the chatbot answers from the data unless you add your own key in its settings.

Pages

Route	Purpose
/	Dashboard: KPIs, top communities, latest registered sales
/communities	Leaderboard with yield, momentum, supply and liquidity grades
/properties	Registered sales, paged, filterable by community and bedrooms
/compare	Up to 3 sales side by side
/map	Leaflet map of the 29 communities (OpenStreetMap coordinates)
/trends	Quarterly median price/sqft, sales volume, same-project momentum
/calculator	Mortgage + short-term-rental calculators
/chat	RAG chat advisor (markdown answers with tables)

Brand assets: /brand/brand-book.html, /brand/luxury-gold.html, /brand/showcase.html

What It Does

Investment Scoring. Each community gets a composite score (20–100) from four factors DLD data actually supports:

Factor	Weight	Measure
Rental yield	40	Per project, median new Ejari rent per m² ÷ median ready price per m²; the community's median across projects
Supply pressure	25	Off-plan sales per ready sale in 2026: units still to be handed over
Price momentum	20	Median change of the same projects from Q1 to Q3 2026
Liquidity	15	Ready sales registered

Each factor is graded A–F and mapped to 1–5 points. Yield bands are absolute (≥7.5% is A); the others were set from the 2026 distribution across communities. Recommendations are bands on that same score: INVEST ≥ 70, HOLD ≥ 50, AVOID < 50. No yield, no score: a community needs at least 3 projects with both new rents and ready sales, and 30 such contracts; otherwise it's "INSUFFICIENT DATA". Momentum needs at least 5 projects sold in both quarters, otherwise it's left out and the other weights are rescaled.

Why per m² and same-project. Ejari rows carry no bedroom count, so rents and sales are compared per square metre, and within the same project: pooling every rent against every sale put International City at 11.6% because a third of its rents came from one newer tower (like for like it's 8.75%). Raw community medians move with which buildings sold (Majan's median rose 52% in two quarters, Marina's 10%), while comparing the same projects shows most communities softened 4–10% over 2026.

Net yield = gross yield × (1 − 8% management − 5% vacancy). It is before service charges, which DLD doesn't publish, and every place it's shown says so.

Data

All from Dubai Land Department open data, January–September 2026, cleaned by src/dld_loader.py:

여기까지가 README 앞부분입니다. 나머지는 원문에서 읽을 수 있습니다. README 전문 보기 ↗