Startup · UAE

Gulf PropTech Marketplace PropTech + GenAI

Built a GenAI property assistant that helps investors compare deals, analyze market trends, and automate tenant inquiries.

Timeline
14 weeks · MVP to production
Team
1 PM, 2 full-stack, 1 ML/LLM engineer, 1 designer
Focus
PropTech + GenAI

The problem

  • Investors were juggling 4+ portals (DLD data, broker listings, spreadsheets, WhatsApp groups) to evaluate a single deal.
  • Agents spent ~60% of their time answering repetitive tenant questions about availability, fees, and documents.
  • No single source of truth for unit status — double bookings and stale listings damaged conversion.

The solution

  • GenAI deal analyst that ingests listings, transaction history, and rental yield data, then ranks opportunities against an investor's thesis.
  • Tenant-facing WhatsApp + web chatbot with retrieval over property docs, contracts, and policies, escalating to a human agent on intent triggers.
  • Unified back-office where agents, investors, and tenants share one timeline per unit — replacing scattered spreadsheets and chat threads.

Architecture highlights

Frontend

React + TanStack Start, server-rendered listing pages for SEO, real-time updates via Supabase channels.

Data layer

Postgres on Supabase with RLS; nightly ETL from public registry feeds + broker APIs; pgvector for semantic search over listings and documents.

AI layer

RAG pipeline (chunked listings + policy docs) on OpenAI + Lovable AI Gateway; deal-scoring agent with tool use for comps, yield, and ROI calculators.

Messaging

WhatsApp Business API + web widget, both routed through one conversation service with human-handoff queue.

Infra

Cloudflare Workers for edge APIs, serverless cron for ingestion, observability via Logflare + Sentry.

TanStack StartSupabasepgvectorOpenAIWhatsApp Business APICloudflare Workers

Measurable outcomes

Indicative figures from the engagement. Exact numbers are placeholders pending client approval to publish.

~3.5×
Faster deal evaluation
From multi-day shortlist to same-session ranked comps.
~62%
Tenant queries auto-resolved
Without agent involvement during pilot.
1
Source of truth per unit
Replacing 4+ disconnected tools.
Working on something similar?

Let's scope a pilot or a discovery sprint.

Start a conversation