carla ramirez works on the hard part of data products: the moment someone has to find one thing in a million rows
sumble 2.0
,
product
,
2026
every question sumble could answer went through one search box and came back as the same undifferentiated table. the brief was "clean up the search page." i argued the problem was that we were modeling three objects as one. i split it into orgs, job posts and people — each with its own view, filter grammar and defaults — and wrote the operator language that now governs every filter surface in the app: one grammar, reused everywhere, instead of a pattern per team. shipped in increments over two months.
sumble 2.0
,
product
,
2026
every question sumble could answer went through one search box and came back as the same undifferentiated table. the brief was "clean up the search page." i argued the problem was that we were modeling three objects as one. i split it into orgs, job posts and people — each with its own view, filter grammar and defaults — and wrote the operator language that now governs every filter surface in the app: one grammar, reused everywhere, instead of a pattern per team. shipped in increments over two months.
marble's design system
,
product
,
2023
i was the only designer at marble, so the system had to make decisions without me. i audited hundreds of legacy components down to a few dozen — the deletions were the actual work. what was left was assembly: patterns a team could stand a new page up from without waiting on design. it outlived me. after i left, about half of all new work shipped from it with no designer in the loop.
MARBLE_DS01.PNG
MARBLE_DS02.PNG
MARBLE_DS02.PNG
MARBLE_DS04.PNG
MARBLE_DS05.PNG
MARBLE_DS06.PNG
marble's design system
,
product
,
2023
i was the only designer at marble, so the system had to make decisions without me. i audited hundreds of legacy components down to a few dozen — the deletions were the actual work. what was left was assembly: patterns a team could stand a new page up from without waiting on design. it outlived me. after i left, about half of all new work shipped from it with no designer in the loop.
MARBLE_DS01.PNG
MARBLE_DS02.PNG
MARBLE_DS02.PNG
MARBLE_DS04.PNG
MARBLE_DS05.PNG
MARBLE_DS06.PNG
bringing price tracking to the zebra
,
product
,
2025
the zebra makes money on quotes. the insurance hub was where people went between them, and it had nothing to do. the ask was engagement; the trap was engagement that competes with conversion. i led a mobile-first, responsive rate tracker where the graph is the engagement and the argument at once. 77% used it, 60% went on to a quote — the second number is the one i was designing for.
Z_REDESIGN_01.PNG
Z_REDESIGN_02.PNG
Z_REDESIGN_03.PNG
Z_REDESIGN_04.PNG
Z_REDESIGN_05.PNG
bringing price tracking to the zebra
,
product
,
2025
the zebra makes money on quotes. the insurance hub was where people went between them, and it had nothing to do. the ask was engagement; the trap was engagement that competes with conversion. i led a mobile-first, responsive rate tracker where the graph is the engagement and the argument at once. 77% used it, 60% went on to a quote — the second number is the one i was designing for.
Z_REDESIGN_01.PNG
Z_REDESIGN_02.PNG
Z_REDESIGN_03.PNG
Z_REDESIGN_04.PNG
Z_REDESIGN_05.PNG
improving sumble's upload flow
,
product
,
2026
import was CSV-only, which quietly meant: to try the product, first go make a spreadsheet. that's not friction, it's activation. i replaced it with paste-and-parse for bulk lists — take the names however you have them, match them, show the match before you commit — then argued it shouldn't live in settings at all. it became the first thing a new user touches.
improving sumble's upload flow
,
product
,
2026
import was CSV-only, which quietly meant: to try the product, first go make a spreadsheet. that's not friction, it's activation. i replaced it with paste-and-parse for bulk lists — take the names however you have them, match them, show the match before you commit — then argued it shouldn't live in settings at all. it became the first thing a new user touches.
marble's payment portal
,
product
,
2024
agents were collecting card numbers over email and phone — a trust and compliance problem being managed as an efficiency complaint. i designed the self-serve checkout and payment portal that replaced them. built on basistheory, which meant the sensitive fields were never ours to style. you can't control the error states you're still accountable for, so most of the work was the frame around the part we didn't own — the frame that decides whether a payment completes or gets abandoned. 90%+ adoption, agent handling time halved.
MP_01.PNG
MP_02.PNG
MP_03.PNG
MP_04.PNG
marble's payment portal
,
product
,
2024
agents were collecting card numbers over email and phone — a trust and compliance problem being managed as an efficiency complaint. i designed the self-serve checkout and payment portal that replaced them. built on basistheory, which meant the sensitive fields were never ours to style. you can't control the error states you're still accountable for, so most of the work was the frame around the part we didn't own — the frame that decides whether a payment completes or gets abandoned. 90%+ adoption, agent handling time halved.
MP_01.PNG
MP_02.PNG
MP_03.PNG
MP_04.PNG
re-imagining honeydew's product landing page
,
product
,
2025
a paid-search landing page where compliance language and scannability fight each other. i rebuilt the hierarchy above the fold and moved trust signals to the exact drop-off points, without softening a clinical claim — responsive, so the mobile fold carried the same argument as the desktop one. a/b tested: 22% higher CVR, 16% lower CAC.
HNYDW_01.PNG
HNYDW_02.PNG
HNYDW_03.PNG
HNYDW_04.PNG
HNYDW_05.PNG
HNYDW_06.PNG
re-imagining honeydew's product landing page
,
product
,
2025
a paid-search landing page where compliance language and scannability fight each other. i rebuilt the hierarchy above the fold and moved trust signals to the exact drop-off points, without softening a clinical claim — responsive, so the mobile fold carried the same argument as the desktop one. a/b tested: 22% higher CVR, 16% lower CAC.
HNYDW_01.PNG
HNYDW_02.PNG
HNYDW_03.PNG
HNYDW_04.PNG
HNYDW_05.PNG
HNYDW_06.PNG