Methodology & Data Sources

Official Dubai Land Department transaction feed (sales, mortgages, grants, lease-to-own). We loaded a 100,000-row stratified sample for the demo - enough to cover every transaction type, area, and year while keeping query latency under ~1s on the demo Postgres tier and AI Gateway token costs predictable. The same pipeline scales to the full 1.5M+ rows; it's a config flag, not a code change.

Historical DLD dataset spanning 1969–2023. Loaded into kaggle_dld_transactions_historic (250K-row sample), kaggle_dld_valuations (73K rows, full), and kaggle_dld_rents (100K-row Ejari sample).

Multi-year bilingual (Arabic + English) DLD transactions dump. Loaded in full into kaggle_dre_transactions via Kaggle API + psql COPY.

UAE rental listings scrape (Bayut/Dubizzle-style). Loaded in full into kaggle_uae_rental_listings - powers rental-yield and price-vs-rent queries alongside the transaction data.

Public league table of 97 Dubai developers with A+ to D grades derived from DLD completion rate, average delivery delay, escrow flags and litigation history. Scraped with Firecrawl on demand, reconciled against DLD project/transaction counts, then re-scored locally using the six-factor RFP weighting (on-time delivery 25% · RERA compliance 20% · units delivered 15% · completed projects 15% · complaint index 15% · ESG 10%). Powers the Developer Scorecard page.

Read-Only SQL Guard

The AI does not write to the database. All queries pass through a run_sql PostgreSQL function with three layers of protection:

1. Regex gate - queries must start with SELECT or WITH.
2. Keyword blacklist - INSERT, UPDATE, DELETE, DROP, ALTER, and similar are blocked.
3. Row cap - results are aggregated through jsonb_agg which caps output at 500 rows per query.

AI-to-SQL Flow

When you ask a question in natural language, the following happens:

1. The AI receives the database schema and your question.
2. It drafts a PostgreSQL SELECT query tailored to the schema.
3. The query is executed via the read-only run_sql function.
4. If the result is numeric or time-series, the AI emits a chart JSON block and the UI renders a Recharts visualization inline.
5. Otherwise, the raw rows are shown in a sortable data table.

Natural-language Q&A over the full DLD dataset. Ask anything in plain English (or Arabic) and get back a SQL-backed answer, chart, or data table.

How it works: Your question is sent to an LLM together with the database schema. The model drafts a read-only PostgreSQL SELECT query, which is validated and executed against the live dataset. Numeric or time-series results are rendered as Recharts visualisations; tabular results appear in a sortable grid.

Interactive rental-listing heatmap covering Dubai, Abu Dhabi, Sharjah and Ajman. Every dot is a real listing coloured by annual rent.

How it works: 73,000+ scraped listings are geocoded and plotted on a Leaflet map. Filters (city, bedrooms, property type, max rent, posting year) are applied server-side so the browser only renders the matching subset. Click any dot for full listing details.

Dubai's most-traded residential projects ranked by transactions, total volume, average price and AED/sqft. Filter by area, year range, and registration type (off-plan vs ready).

How it works: Aggregates 1M+ DLD sales transactions at the project level. A top-project card highlights the #1 most-traded tower. The table supports sorting by volume, price, or transaction count, with inline bar visualisations for quick comparison.

Rental premium calculator for every Dubai Metro station. See how much extra tenants pay to live within walking distance of each Red and Green line stop.

How it works: Rental listings are grouped by proximity to each metro station (radius configurable from 800m to 4km). The station average AED/sqft is compared against the citywide average from 34,000+ listings. Stations are ranked by premium percentage.

Off-plan vs ready market gauge. Track the speculative cycle by measuring off-plan transaction share, volume, and price premium over time.

How it works: Splits DLD project-level sales into off-plan (developer primary sales) and ready (secondary resale). Calculates share by transaction count and by volume, plus average AED/sqft for each category. Hot areas (highest off-plan share) and cool areas (ready-heavy) are ranked.

Valuation gap detector — compare DLD-registered bank valuations against realised sale prices to detect bubbles and financing discounts.

How it works: For each area, computes the median bank valuation AED/sqft and the median realised sale AED/sqft. The gap percentage = (sale − valuation) / valuation. Positive gaps signal market prices running above bank appraisals; negative gaps suggest financing-side conservatism.

New vs renewal rent comparison. Measures how much newly signed Ejari contracts diverge from renewals — a proxy for tenant repricing risk and landlord pricing power.

How it works: Ejari rental records are classified as new (first-time registration) or renewal (re-registered lease). Median AED/sqft is computed for each group by area and year. A large new > renew gap means sitting tenants are under-market and face sharp increases on renewal.

Year-by-month transaction heatmap for sales and Ejari rentals. Spot quiet months, peak seasons, and long-term shifting patterns.

How it works: Transactions are binned by year and month, then normalised into a heatmap. Click any cell to drill into month-level details: transaction count, median/mean/total value, top areas, top property types, and top projects for that month.

Developer reliability scorecard — one A+ to D grade per Dubai developer, computed from six weighted criteria straight out of the IRETP RFP.

How it works: WealthIQ's public league table is scraped via Firecrawl and reconciled against DLD project and transaction counts. Each developer is re-scored locally with the six-factor RFP weighting: on-time delivery 25%, RERA compliance 20%, units delivered 15%, completed projects 15%, complaint index 15%, and ESG certification coverage 10%. Results are cached in the developers table and refreshed on demand from the page.