04·AI / ML
Data Atlas
An entity-centric valuation platform and on-demand AVM over 538K entities across 28 sources, with a MAPE and RMSE calibration and backtest harness; an internal snapshot showed ~7% MAPE.
ResultTurns multi-day analyst cycles into an on-demand AVM brief, ranked opportunities, and an accuracy snapshot the valuation team can audit.
Discover
Why it exists.
BARNES Dubai's valuation team was running multi-day analyst cycles to produce ranked opportunity briefs, pull RTA transactions, cross with KHDA, eyeball portal scrapes, layer in flood and vegetation risk. Each brief was excellent and was finished a week after the listing moved.
Atlas collapses that cycle into an on-demand surface. An entity-centric platform: 538K entities reconciled across 28 DLD, Barnes-internal, and geo sources. The customer-facing AVM Tool sits on top, a tab the analyst opens, types an address into, and reads back a valuation with confidence bands, comparables from the Dubai Land Department, a six-month price forecast, and a six-dimension investment score.
The hard part is not the FastAPI service; it is the reconciliation. Twenty-eight DLD, Barnes-internal, and geo sources disagree on what a building is called, how a unit is referenced, and which currency a price is in. Atlas's job is to make them agree without lying, and to surface the AVM's accuracy (an internal snapshot showed roughly 7% MAPE) so the team can audit it.
StackPython · FastAPI · PostgreSQL · Cesium · React · Machine-learning pipelines
Define
The brief.
Collapse the valuation team's multi-day analyst brief into an on-demand surface, without averaging away the disagreements between twenty-eight sources.
Develop
What it took
Skills behind it.
Primary discipline plus the support stack.
Skills demonstrated
- Backend EngineeringPrimary
FastAPI service on PostgreSQL with hex-grid spatial index, confidence-tiered entity ontology, and entity resolution across 538K entities from 28 sources.
- AI & Machine LearningSecondary
Automated valuation model with a MAPE, RMSE and calibration backtest harness (an internal snapshot showed roughly 7% MAPE), six-month community-calibrated forecasts, and a six-dimension risk-adjusted investment score.
- Data PipelinesSecondary
28 sources reconciled into one analyst surface, from DLD transactions / buildings / brokers / projects / developers / permits / rent / freehold / land to Barnes Internal LLM market, Dubai-coordinate geo, and more. Real Dubai Land Department transaction data, zero mock.
- Frontend EngineeringSecondary
React + Cesium AVM tool + DLD-comparables panel + accuracy snapshot dashboard. The customer-facing valuation surface the team opens daily.
Develop · Build
How it came together.
- 01
Confidence tiers everywhere, not just on prices
Every entity carries a confidence tier, high for direct DLD transaction records, lower for inferred matches across portals, lowest for derived calculations. The AVM Tool surfaces these tiers explicitly: an analyst should be able to tell, at a glance, which numbers came from the registry and which were inferred. Hides nothing, claims nothing it cannot back.
AI & Machine Learning - 02
AVM with a published accuracy snapshot, not a black box
The valuation model publishes its own running backtest: 12-week rolling MAPE per community, calibration tracked predicted-vs-realised, top-six communities ranked by sharpness (Dubai Marina 5.81%, Palm Jumeirah 6.18%, Dubai Hills 6.42%, Downtown 6.94%, The Springs 7.04%, Business Bay 7.12%). The team can see exactly where the model is sharp and where it is soft before they trust it for a brief.
AI & Machine Learning - 03
Hex-grid geospatial risk, not polygon overlays
Flood, vegetation, urban classification, and walkability were modelled on a hex grid rather than on irregular polygons. Hex cells aggregate faster, query consistently, and visualise cleanly at multiple zoom levels. The same hex layer renders in Cesium for the analyst surface and powers the six-month forecast pipeline downstream.
Data Pipelines - 04
Six-month forecasts conditioned on local supply
Forecasts are community-calibrated, not free-floating extrapolations, each is conditioned on local supply (off-plan launches, completion timetables) and macroeconomic anchors (EIBOR, World Bank). The model surfaces both the prediction and the contributing factors, so the analyst can challenge the forecast on the factor that looks wrong.
AI & Machine Learning - 05
AVM Tool that reads like an analyst memo
Built the React + Cesium AVM surface to read like the memo the analyst would have produced manually, narrative paragraphs, ranked opportunity list, DLD-comparable transactions, map, supporting tables. The platform's job is to draft the memo; the analyst edits and signs. No dashboards full of widgets nobody clicks.
Frontend Engineering
Develop · Forks
Decisions on the record.
The few calls worth defending. Each one is a fork; the other branch would have been a different project.
Decision · 01
Why Cesium for the map layer, not deck.gl
Cesium is georeferenced from frame one and shares its coordinate system with the Splat Viewer city-scale overlay, so a building rendered in the AVM tool can hand off to the 3D scan without a re-projection. deck.gl is faster to start with but the cross-product handoff would have cost more later than Cesium's heavier upfront cost.
Decision · 02
Why a confidence-tiered entity ontology, not raw connectors
Twenty-eight DLD + Barnes-internal + geo sources disagree on the same entity all the time. Resolving them into a tier-scored ontology (PRIMARY vs. CORROBORATING vs. SUPPLEMENTARY) makes the disagreement visible instead of averaging it away. The AVM tool surfaces the tier on every claim so valuers can audit the chain, not just the number.
Decision · 03
Why a published accuracy snapshot, not a hidden one
The backtest accuracy is designed to sit in the tool's own UI, not a hidden internal report, so the team that opens the tool sees the same number they would show anyone else. The principle is that the credibility cost of hiding accuracy beats the embarrassment cost of a number you have to defend.
Deliver
What shipped.
Turns multi-day analyst cycles into an on-demand AVM brief, ranked opportunities, and an accuracy snapshot the valuation team can audit.
In production with the BARNES Dubai valuation team. The AVM Tool replaces that manual cycle with an on-demand brief; the analyst spends time on the harder cases instead of repeating the same lookup five times an hour. The data pipeline backbone is also the source for the Barnes Dubai LLM multi-agent research system.
By the numbers
MAPE
Internal backtest snapshot
Entities
Data sources
Frames
Selected from the work.
3 frames
01Overview 02Comparables 03Transactions