YieldX: Dubai transaction exports turned into building valuations

01 / Overview

About YieldX

02 / The challenge

What YieldX needed to solve

The data was messy, the comparisons had to be fair, and every figure had to be honest about how much evidence stood behind it.

1. 01

   Transaction exports arrive as large Excel files that overlap with each other, so the same sale can show up more than once.

2. 02

   The exports mix apartments with villas and hotel units, which cannot be valued against each other.

3. 03

   A flat had to be compared first with its own tower, then sister towers, then similar buildings nearby, with recent sales counting for more.

4. 04

   Some buildings have many sales and some have only two, so every figure needed a clear label for the evidence behind it.

5. 05

   Runs are long and the hosted database can drop idle or mid-query connections, so the pipeline had to survive broken links.

03 / What we built

How we built YieldX

We built a FastAPI service and a Python valuation engine on Supabase Postgres, with a one-page Next.js panel in front for the operator.

• 01

  One-page bulk upload

  The operator drops every sales and rental Excel file at once. There is no need to sort or rename files before uploading.

• 02

  Automatic file detection

  Each file is identified as sales or rental from its header row. The panel shows how many of each were found before the run starts.

• 03

  Duplicate-safe loading

  New rows land in a temporary table first, and only rows not already stored are inserted. Uploading the same file twice adds nothing, and the panel reports how many duplicates were skipped.

• 04

  Seven-step valuation engine

  The engine filters outliers, groups sister towers and weighs comparables before writing results. Each building and bedroom type gets a valuation built the same way every time.

• 05

  Three comparable tiers

  A unit is compared with its own tower first, then sister towers, then similar buildings nearby, with recent sales weighted more. This keeps comparisons fair across very different buildings.

• 06

  Confidence labels

  Each result is labeled high, medium or low confidence, with plain-text flags such as how many direct comparables stand behind it. Analysts can see at a glance which figures to trust.

• 07

  Step-by-step run tracking

  The panel shows each stage: uploading files, saving raw data, running the engine and complete. Only one run is allowed at a time, because each run rewrites thirteen output tables.

• 08

  Connection retries

  The engine checks the database connection and retries when the hosted database drops a link. Long runs finish instead of failing halfway.

04 / The results

What changed for YieldX

An operator loads a batch of Excel exports and the system writes building-level valuation records into tables ready for a dashboard.

• Each record carries a low, middle and high fair value plus rental yield fields.
• Plain-text flags show how many direct comparables support each value, so uncertainty is stated, not hidden.
• Re-uploading overlapping exports no longer inflates the data, because duplicates are skipped automatically.
• Operators run the full process from one page without touching the database or scripts.
• The API runs on Render and the panel on Vercel, so the tool is available without local setup.