Portfolio  /  Case studies  /  Property Health Monitor

✦ AI case study

Property Health Monitor  ·  Porch Moving Group  ·  6 min read

Catching broken lead data in 48 hours, not 15 days

Same failure, caught about 87% faster.

Every lifecycle workflow depends on clean contact data. When a partner integration or a form breaks, contacts arrive with empty fields, or stop arriving, and nothing looks wrong. I built a monitor that checks every lead source against its own normal.

✦ Try the live demo

IMAGE · dashboard (phm-dashboard@2x.png, waiting for upload)

FIG. 01   The heatmap and trend on demo data. Marketplace is down 40% and has gone quiet for 27 hours.

My role

Lifecycle Marketing Manager (designer and builder)

Focus

Data Quality, Lead Source Monitoring, Marketing Operations

Tools

Apps Script, Claude, HubSpot, React, Sheets, Slack, Gmail

Scope

The full HubSpot portal, across 27 lead sources

Year

2026

The 30-second version

The problem

What was going wrong?

A lead source once stopped sending contacts, and it took 15 days before anyone noticed. Broken forms and integrations hit one source at a time, and program-wide numbers hid them.

What I built

What did I make?

A monitor that checks every lead source against its own normal and flags missing fields, volume drops, and sources that go quiet. Claude explains what to check first, and a live dashboard shows it all at a glance.

The impact

What changed?

The next time a partner stopped sending leads, it caught the drop in under 48 hours (about 87% faster), and we alerted the partner before any downstream impact. It grew from 142 key properties to the full HubSpot portal.

Skills shown

What it shows

Data quality

Marketing ops

AI tooling

Dashboard design

Data visualization

React

Results

<48 hrs

The catch that proved it

Same failure. This time, under 48 hours.

Before the monitor, a lead source stopped sending leads and we only noticed when a 15-day workflow finished with no new enrollments. A few months later, a partner stopped sending leads again. This time the monitor flagged the drop within 48 hours, and we alerted the partner before it had any downstream impact.

Faster to catch a lead source outage

~87%

From 15 days to under 48 hours

Contact properties covered

848

The full HubSpot portal, across 27 lead sources. Version 1 checked 142.

Kinds of breaks it catches

3

kinds

Fields that stop filling in, volume drops or spikes, and sources gone quiet

Context & challenge

In short: a broken form or integration hits one lead source at a time, and program-wide numbers hide it.

Our workflows segment and branch on HubSpot properties, and leads arrive from our own sites and a dozen partners. A broken form or integration usually hits one source at a time, and program-wide numbers hide it. Workflows keep running; they just enroll fewer people, or send them down the wrong branch because a field is empty.

Before the monitor, a lead source stopped sending leads and we only noticed when a 15-day workflow finished with no new enrollments.

The challenge was to

Within a day or two

Spot a broken source within a day or two, not weeks.

Real breaks, not noise

Separate real breaks from normal variation, especially for small, low-volume sources.

What to check first

Point the team to what to check first.

The whole portal

Cover the whole portal without pulling every contact every time.

My approach

In short: every lead source is judged against its own normal, and severity is kept separate from the size of a change.

The monitor compares the last 3, 7, or 14 complete days with the prior 30 or 60 for every lead source, and uses standard deviations from that source’s daily norm, not a fixed percentage.

Check

Warning

Critical

Field fill rate

Warning

1.5 std. dev. below normal and at least 5 points down

Critical

2 std. dev. below, or dropped to about 0%

Contact volume

Warning

Unusual drop or spike of at least 20%

Critical

A regular source drops to zero

Gone quiet

Warning

Silent for 4x its usual gap

Critical

No contacts in 24 hours from a source that sends every day

Each source is judged against its own daily norm, so a partner that sends a handful of leads a day and one that sends hundreds never share a threshold.

Field importance

Mark fields Critical, Tracked, or Ignored. Critical fields (name, email, phone, brand, move date, order status) escalate sooner.

Noise thresholds

Sources with too few contacts, and fields a source rarely fills, aren’t judged, so small partners don’t raise false alarms.

New vs. ongoing

Every flag shows how many days in a row it has been open.

Known issues

Mark an issue as expected with a note, and it stops alerting until the numbers change.

Claude explains it

Groups the flagged fields by likely cause and says what to check first.

The dashboard

In short: I designed it to be read in seconds, using data visualization practice.

Heatmap hero

The main view is a field-by-source heatmap on a colorblind-safe scale (the image at the top of the page).

Two separate signals

Severity is marked with a ring or a notch, separate from color, so “big change” and “statistically unusual” are never confused.

Sentence headline

The top of the page states the finding in one sentence instead of a big number.

Trend chart

Daily dots sized by sample, with a typical-range band.

Biggest drops

A dumbbell chart ranks the largest before-and-after drops.

The build

In short: version 1 ran in Apps Script and Slack; I rebuilt it with Claude as a live dashboard on the whole HubSpot portal.

Version 1

Apps Script + Slack

•  Five scheduled jobs, twice a week, later daily

•  142 key properties across 30 sources

•  Posted to a Google Sheet, Slack, and email

Version 2

A live dashboard, rebuilt with Claude

•  The full HubSpot portal: 848 contact properties across 27 sources

•  Three views: Monitor, Sources, Fields

•  Samples each day and caches finished days, so a refresh only fetches what’s new

•  Drafts one alert email in Gmail for every brand, and never sends on its own

Along the way I

Carried over the best of version 1: the lead source grouping, the critical field list, volume alerts, and the 3-day window.

Raised thresholds after testing on real volume, so statistically real but tiny shifts at high volume don’t alert.

Debugged the live HubSpot connection and documented how it actually behaves, with a recovery guide so anyone can rebuild it.

Version 1 · Slack alert · May 26, 2026

⚠️ Property health alert

PODS partner

Phone numbers dropped to 0% on recent contacts

HireAHelper.com

The lead source field stopped filling in

It also flags rising fill rates and volume spikes, which often point to a pipeline change.

Rebuilt from the real alert; counts and links removed.

First live run · Sep 23, 2026

18 fields dropped at once

It flagged one source, the HireAHelper unified checkout. The cause was expected behavior, not a break: those contacts move from unknown to booked to completed, so booked-stage fields empty out as jobs complete. That’s exactly what Mark as expected is for: note the reason once, and the monitor stops raising it.

Who reads it

My team and marketing leadership (our head of marketing and CMO) read the flags, and sometimes product leadership too, when a fix needs a form or an integration changed.

Try the demo

In short: click through the real dashboard on synthetic data, and ask it for likely causes.

DEMO · iframe goes here (HTML widget, added by hand). Fallback image: phm-demo-top@2x.png, waiting for upload

Live demo

Interactive demo with synthetic data. Click any cell, tab, or filter, change the date windows, and run “Likely causes”. Open full screen →

✦ Open the demo full screen

What I learned

Speed is the whole point

48 hours vs. 15 days is the difference between a quick fix and lost revenue.

Compare each source to its own normal

A partner that sends a handful of leads a day and one that sends hundreds can’t share a threshold.

Design for trust

Keeping severity separate from the size of a change, and letting people mark known issues, keeps an alert tool from being ignored.

Expected changes aren’t bugs

The first live flag was a normal status change; the tool needs a way to learn that once.

What’s next

In short: make it run on its own, and check values as well as whether fields are filled.

Scheduled alert email

A daily alert email that runs without the dashboard open.

Value checks

Catch move dates in the past, malformed ZIP codes, and placeholder text.

Sub-source drill-down

Break flags down by sub-source, and link any flagged cell to example contacts in HubSpot.