PropTech / property management · 2024 — 2025
Xprop (大廈通)
AI property management for Hong Kong buildings — WhatsApp chatbot + management portal.
- >30%
- lower pricing than incumbents
- 50,000
- buildings in the Hong Kong market
- 10%
- penetration target used for sizing
- 3
- property players anchored for validation
01About
What it is
Xprop (大廈通) is an AI layer over the paper-heavy world of Hong Kong building management. Tenants talk to it on WhatsApp — Cantonese voice notes and photos included. Management sees everything in one portal.
The status quo
Today this runs on group chats, filing cabinets, and manual re-typing. Xprop turns the messy inbound into structured, routed, searchable tickets and records.
The wedge
The wedge: Cantonese-first, WhatsApp-native, and priced well below incumbents — validated with three property players, anchored on JLL, proven in a live building pilot.
02How it works
Inbound — issues & requests
01
WhatsApp inbound
Voice notes · photos · documents
02
ASR + document vision
Cantonese speech and scanned bills → data
03
Classify & route
Tickets, logs, categories
04
Management portal
Dashboards by building, priority, status
Outbound — documents & notices
01
Management upload
Bills and notices, added to the portal
02
Auto-tag & assign
Matched to units against the tenant list
03
HITL approve
A person confirms before it publishes
04
WhatsApp delivery
Tenants see it land in the same chat
Per-tenant visibility — Each tenant only ever sees documents tagged to their own unit
03My role
The product where I was most end-to-end — the hands-on PM and designer, in between biz and tech.
The PM half
- Ran client discovery with JLL; turned front-line pains into use cases and user stories
- Owned GTM and the sales narrative — proposal deck and flyer included
- Ran a phased, gated delivery plan across a 90-item estimated backlog
- Set up the Meta / WhatsApp Business accounts and attached the live channel myself
The UX/UI half
- Designed the logo and full brand — flyers, banners, print
- Wireframed the portal, mobile screens, and dashboard flows in Figma
- Shaped the bilingual chatbot menus and wording
- Made the demo-day calls: hide the god-view, show the latest bill only, red helper text
04Use case
Tenant > Staff > Building Owner
- 01
A tenant sends a Cantonese voice note on WhatsApp.
- 02
ASR transcribes it. System classifies it and routes to management portal.
- 03
Property staff sees it on the portal dashboard, and replies.
- 04
Building Owner see all issues via dashboads and alerts.
05Challenges & solutions
Cantonese speech recognition stumbled on property trade jargon — the abbreviations and site terms staff actually speak into voice notes (批灰, 收口, MIC).
Shipped voice-to-text with human-in-the-loop correction rather than sinking pre-sales dev effort into fine-tuning. Logged the jargon as a labeling need for a later pass — enough accuracy to keep demos credible, without over-investing before the deal was real.
Property managers have to justify any new spend internally, and tenant budgets get scrutinised sharply past a certain line — price it wrong and adoption dies before the pilot.
Set a flat per-building rate that lands under that scrutiny line, keeping the buying decision a fast yes instead of a budget fight. Pricing became an adoption lever, not just a revenue figure.
The product-lead seat sat between a commercial side that wanted signed deals now and an engineering side with finite bandwidth — the familiar no-budget-no-build standoff.
Owned that tension instead of dodging it, using demos and proof-of-concepts to de-risk pre-sales before spending real dev time. Bridging the two sides was the job — not a detour from it.
06UX / UI
Tenant side
The tenant side is WhatsApp — no app to install, nothing to learn. Menu-first flows that survive being used one-handed in a lobby.
Management side
The management side is a portal command center — per-building chat, ticket detail, dashboards by priority and status — designed in Figma, skinned on an open-source PM tool.
Details that landed
Small calls made the demos land: defaulting to one building, simplifying bill retrieval, renaming log IDs after customer feedback.
07System decisions
- WhatsApp instead of an app — tenants already live there. The single biggest adoption decision.
- On-prem storage to win trust with nervous building owners.
- Menu-driven, not open-ended chat — a predictable path beats an impressive-but-flaky one.
- Low-confidence AI extractions route to a human. A wrong number on a bill is worse than a slow one.
08Workflow tools
Figma — Wireframes, UI screens, and dashboard flows.
Framer — A shareable proof-of-concept site to sell the concept fast.
Linear — The PM platform for roadmaps, milestones, and issues.
Datadog — Per-meeting cost measurement feeding the pricing model.
WhatsApp Business API (Woztell) — The tenant channel — chat, broadcast, numbers.
Otter.ai — Discovery calls transcribed into searchable, citable evidence.
Mermaid — The architecture and data-model diagrams, authored as code.
AI infographic HTMLs — Explainer one-pagers generated as HTML for client conversations.
09Data
The raw material
Xprop runs on the genuinely messy documents a building produces — notices, bills, permits — arriving as photos, voice notes, and scans.
Sources
- WhatsApp submissions: text, Cantonese/English voice notes, photos, documents
- Building notices, bills, and billing documents, per unit
- Structured master data: building codes, block/floor/unit
- Real pilot documents from a live building
Where it sits
Incumbents are pricey Western PropTech that can't do Cantonese. Xprop sits in the opposite corner: Cantonese-first, WhatsApp-native, tuned to the documents HK buildings actually generate.
10AI
- Cantonese speech recognition via Whisper — the main incumbent simply can't do it.
- Document vision pairs invoice and contract values; low-confidence goes to a human.
- On-prem LLM plus Azure OpenAI for extraction and classification — on-prem where privacy demands it.
- I wrote the production billing-extraction prompt myself — a structured JSON schema keyed on each payment-advice number.
- Honest scope: open-ended semantic search was flagged as an over-promise risk and kept out.
Whisper
Azure OpenAI