Pro Service
Leo Ramos
Oct 22, 2026 8:20 AM
Called Tasha: she's home from 1 PM. Boro Heating has the lockbox code and will text when they're on the way.
As the only product designer, I led design for the platform 200+ employees in the US, Argentina and Peru use to run Belong: payments, maintenance, leasing, inspections and admin. I worked with 3 squads, 3 PMs, engineering leads and the VP of Product, from the design system to Phin, its AI layer.
Foundation member profile for a homeowner. On the left, the member card with persona tags, contact details, sentiment score, NPS and CSAT, then their home and team. On the right, the Summary tab: Phin's summary, team tasks, appointments, the home, the plan and the team.
Phin summary
What matters for this member right now, so nobody has to read ten tabs to catch up.
Phin summary
What matters for this member right now, so nobody has to read ten tabs to catch up.
Persona tags
Their cash flow, investor type and distance from the home, so every conversation starts in the right place.
Sentiment, NPS and CSAT
How this person feels, visible to everyone who talks to them.
Team tasks and appointments
What's open on this member's homes, how overdue it is and who's coming next.
Home, plan and team
The home, the agreement and the assigned team, without leaving the page.
Phin summary
What matters for this member right now, so nobody has to read ten tabs to catch up.
Everyone was reacting.
More than 200 people kept about 2,500 rental homes running, on tools that had grown piece by piece.
Work arrived from everywhere and nothing said what mattered most. People spent the day deciding what to do, and supervisors only saw problems once they escalated.
“How do we make 200 people proactive, focused on the customer and not the tool?”
One designer, three squads.
I couldn't design every screen for three squads, so the system had to do most of the work.
We started from shadcn and React and adapted them for dense operational work. Because it was already code, engineers adopted it fast.
Live components
Persona tags
Who you're talking to, at a glance: one tag per dimension, and a tooltip says why.
Task card
If you can triage it from the top two layers, the hierarchy is right.
Buttons
One size, 40px, so density is never a choice. One green button per view.
Sentiment, NPS and CSAT
How this relationship is going, on the member page.
Input and select
The placeholder floats up into the label, so a filled form still names each field.
Phin summary bubble
The AI speaks as a character and stays short until asked.
Tabs
Two weights of the same control: a quiet underline and a filled version.
Tags
Twenty colors. The border is the color and the fill is 10% of it.
SLA indicator
On a task, three colors answer one question: how late is this?
Selection controls
Each kind of choice has one control.
Member
One way to show a person, with or without a photo.
Tooltip, dark
For answers that take more than a line, as on persona tags.
Button group
A single choice with every option in view, in one tap.
Rating input
One component, with stars for quality and dollars for cost.
Once the components existed, I began building the front end myself with AI tools. Two-week sprints that released in week three became weekly releases.
One platform, many jobs.
Internal tools usually get whatever is fastest to build. I argued the opposite: their experience becomes the customer's experience.
Foundation opens on a live map of each person's homes, with layers for vacancies, repairs and renewals.
An AI-ranked work queue with live team metrics, in a version for each role.
Design ruleThe tool decides, the person acts.
One page for every home and person. It changes with the home's stage, so the right information stays on top.
Design ruleEverything in one place, clear in 30 seconds.
Every repair, from the resident's report to the last photo, on one page.
Design ruleMake performance visible.
Inspections by room and feature, with photos and a recommendation, linked to the home.
Design ruleEverything in one place, clear in 30 seconds.
From ranking the work to doing it.
The AI came from a design question: what would it take for people to stop deciding, defining and searching? Belong's assistant, Phin, grew in four steps.
AI ranks the Control Panel by urgency, SLA due dates and time since last touch. People start at the top instead of sorting.
4 Tasks
Order By: Highest Priority
No Heat Emergency Repair
Last Updated Oct 22, 8:15 AM EDT
1 Day Overdue
Leak Under Kitchen Sink Repair
Last Updated Oct 22, 7:31 AM EDT
Recent Move-In Due Today
Dryer Vent Repair
Last Updated Oct 21, 5:40 PM EDT
Due In 4 Days
The Phin summary puts what matters on a page into one or two lines. Moments That Matter reads every call, email and payment, and surfaces the few that shaped a member's experience.
Tasha GreenResident
19 Juniper Way, Bed-Stuy, NY

Tasha has lived at 19 Juniper Way since March 2025 and has never missed rent. She reported no heat yesterday and has followed up twice; Boro Heating is booked for 2–4 PM today. She rated her last repair 5 stars and plans to renew in the spring.
Moments That MatterBeta
Built on Claude, Phin chat knows member data, history and internal guides, and answers about the page you're on.
Ask me anything you want about this home.
From chat, people can create pro service requests, handle authorizations and pull knowledge. Phin drafts; a person confirms.
What can Phin help you with today?
Team leads curate the quick prompts their people see, so Phin starts from real jobs: refunds, failed payouts, vendor no-shows.
The hardest part wasn't the interface.
The product changed how people worked and how they were measured. Its metrics drove performance and bonuses, which caused pushback early on.
Adoption came through feedback, training and product changes, until working proactively was simply the job.
Oct 2024
Launch
Foundation replaces the old admin.
Friction
Pushback
A new tool and a new way of being measured, at once.
Iteration
Listen and adjust
Feedback, training and product changes, release after release.
Today
Proactive by default
Just how the work gets done.
Along the way, response times went from 3–5 days to 1–2 days to almost immediate, customer satisfaction rose from 5–6 to 8–9 out of 10, and releases went from every three weeks to every week, with design shipping the front end.