PropTech / property management ·

Xprop (大廈通)

Turning a building's paper trail into a product four kinds of people can use.

3
Key property management market players anchored for Xprop's go-to-market
~5
Buildings live in pilot — against a 4,000-building prospect pool
>700
Users across the pilot buildings
Xprop's WhatsApp-native building management interface

01

Product Overview

Shipped into one real building, on real tenant data — not a deck.

Role
Product Manager + Designer
Timeline
2024.10 — present
Platform
WhatsApp · web portal
Scope
Discovery through live pilot

Xprop (大廈通) is an digital layer over the manual-heavy world of Hong Kong building management. Tenants talk to it on WhatsApp, leave Cantonese voice notes and photos included, and management sees everything in one portal. I ran the discovery, designed the solutions and ship it to biz team.

Xprop marketing landing page
The public face of the product — how the pilot was pitched to building management.

02

4 user segments

Four user types in one reporting chain and each has different interests.

  • A tenant reporting a leak by phone in a building lobby

    Reports without an app

    Pays for two buildings — home and office — with no software for either.

    Actions

    • Report building issues
    • Track bills and notices
  • A front-line building officer on a handheld radio at the lobby counter

    Works out which unit it is

    Runs the day on the ground, working out which unit and block an issue belongs to.

    Actions

    • Coordinate service, log cost and actions
    • Report to the property manager
  • A property manager reviewing building performance charts on a meeting-room screen

    Answers for the cost

    Owns the management budget and the service the building is judged on.

    Actions

    • Reporting on cost and operations
    • Task delegations
    • Report to the building owners
  • A building owner looking out over the Hong Kong skyline from an office tower

    Answers for the reputation

    Owns the buildings — an asset and a reputation, not a maintenance queue.

    Actions

    • Source solutions to get revenue
    • PR crisis prevention
    • Portfolio for connections

03

The Challenges that we face

A wrong bill ID logged once becomes a short payment nobody can explain.

I ran discovery workshops with all 4 personas and asked them to bring what they actually work with. Sessions surfaced the real data: notices, bills, permits, work orders — on paper — plus a 100-person WhatsApp group where the same issue got reported multiple times over.

Xprop property document capture
Building document (bill)
Xprop property document capture, second example
Building document (notice)

Discovery was client-led: near-weekly calls with the client's portfolio lead from October 2024, run through Otter.ai so every decision traced to a timestamp. I asked for the buildings' own paper rather than a requirements doc.

  1. A tenant reporting a leak by phone in a building lobby

    Tenant

    Sends a Cantonese voice note — water leaking, 8/F lift lobby.

    • a voice note
    • a photo
    • the unit number, in their head
  2. A front-line building officer on a handheld radio at the lobby counter

    Front-line officer

    Works out the block and the unit, calls a third-party service, sees it fixed.

    • the contractor's quote
    • what was actually done
    • what it cost
  3. A property manager reviewing building performance charts on a meeting-room screen

    Property manager

    Cannot see that an officer logged the wrong bill ID on purpose — the charge lands against the wrong record, and the tenants never pay it in full.

    • a bill ID nobody re-checked
    • short payments, no explanation
  4. A building owner looking out over the Hong Kong skyline from an office tower

    Building owner

    Lost HK$1M on a tenant because payment receipt were mis-recorded long ago(Lack of documentation system and logs for legal to take on)

    • HK$1M lost
    • a receipt nobody can produce
    • a case legal will not touch
Workshop whiteboard mapping the Xprop pipeline — ASR to text, then log type, unit and block, location and time
The working session behind the record — speech to text, then log type, unit and block, location and time, and the question that settled the data model: which of these can a rule fill, and which need a model?
Discovery documents & data model

04

WhatsApp in front, Dashboard at the back

Meet people where they already are, then do the structuring behind the glass.

The end-to-end journey, as shipped

  1. 1

    A tenant sends a Cantonese voice note on WhatsApp.

  2. 2

    Speech recognition transcribes it; the system classifies and routes it.

  3. 3

    Front-line staff pick it up on the dashboard and reply in-thread.

  4. 4

    Manager and owner watch it close on dashboards and alerts.

Xprop WhatsApp bot on a phone

One record

TKT-204 · 漏水 — 8/F lift lobby

狀態
In progress
單位
8/F lobby
優先級
High

Tenant

WhatsApp

writes · reads own unit
  • Reports an issue by voice note or photo — no app, no login, one hand.
  • Bills and building notices, filtered to the unit they pay for.

Front-line officer

Portal ticket

writes at source
  • One thread carrying both the tenant and the third-party service.
  • Cost and action logged onto the ticket while doing the job, not after.

Property manager

Dashboard

reads · delegates
  • Cost and operations roll up on their own — no monthly rebuild.
  • Delegation by assignee, with the response clock already running.

Building owner

Digest & alerts

reads only
  • The portfolio across buildings, without the ticket detail.
  • An early warning while an issue is still an issue, not a PR problem.
Xprop issue detail screen showing a ticket's full historyXprop analytics dashboard showing cost and response time across buildings

The record the officer fills in while the leak is being fixed is the same one the manager reports on and the owner is warned by. Nobody upward re-types it — that is the product.

Loading diagram: How a voice note or photo becomes an auditable record.…
How a voice note or photo becomes an auditable record.

Deterministic

Bill extraction

PM-authored rules wrap the model — dedupe by PAYMENT ADVICE NO., valid JSON only.

Checked, not trusted

Added human-in-the-loop review and a VLM confidence score, so every extracted field is verified rather than assumed correct.

Gray zone

Who signs off

Routing is part rule, part model classification — and "who carries responsibility" for a PMSA call was never closed.

Manager's call

Taxonomy set by hand; delegation left with the manager. The system proposes, a person still decides.

Probabilistic

Cantonese jargon

Local building jargon is a named weak spot in ASR, and no WER benchmark exists for it.

Correction loop

Shipped on human correction — low-confidence output is flagged for review instead of claiming an accuracy number nothing backs.

Fixed rule, model's best guess, or still a person's call.

05

Results and Impact

Pricing is the strongest commercial call in this case. Scale is the honest miss — worth saying plainly, not softening.

3
Property management players anchored

Key market players secured for Xprop's go-to-market.

~5
Buildings live in pilot

Against a 4,000-building prospect pool.

>700
Users across the pilot buildings
  1. 01

    Per-head

    Scales with headcountper person

    Rejected

    Grew with the number of residents, so it punished exactly the biggest buildings the pipeline needed most.

  2. 02

    Tiered A / B / C

    Three bands to choose fromper building

    Rejected

    Three rates instead of one meant every deal became a negotiation over which tier applied before it could be about the product.

  3. 03

    Flat, per building

    One rate, agreed per portfolioper building

    Shipped

    One number, no tier to argue over — and it's set with the client against their own budget line rather than read off a list price, so a portfolio can start small and widen without repricing.

The client's portfolio lead set the real constraint: there's a spend line below which nobody asks questions and above which everyone does. The rate was placed on the safe side of that line — their line, not a cost-plus spreadsheet's.

Five months, six pricing models tried, three shapes shown here. Figures stay out of it on purpose: what shipped is a flat rate agreed per portfolio, so the shape is the transferable part, not the amount.

06

AI workflow at a glance

AI did the reading and the routing; every price, compliance and sequencing call stayed with a person, by design.

  • Vision + ASR — Prompt-engineered by hand — reads bills and Cantonese voice notes into structured records.

  • Mermaid — Architecture and the data-model class diagram authored as code.

  • Framer — A shareable proof-of-concept site, so the idea sold before it was built.

Teammates Involved

Agents did the typing. These are the people who set the direction, argued the trade-offs and reviewed what shipped.

Tech leadTech lead
Business leadBusiness lead
Client's portfolio leadClient's portfolio lead
Client's technical managerClient's technical manager