On July 22, I started deleting my own design
Connection Central is where Quickbase admins work out which systems, connections, tables, and pipelines are touching company data. The first version, designed by Johnny Lee, already exposed a lot of information. My v2 work in 2024 was supposed to make the relationships easier to reason about.
Research gave me a better information architecture and, naturally, enough good ideas to miss the release date. On July 22, I removed the richer accordion and popover detail from the MVP because I had designed more than engineering could responsibly build before GA. The later July 20, 2025 through July 19, 2026 read-back records 3,110 accounts using the page: real reach, but not proof that the redesign made a difficult investigation easier.
My role
- Role: Lead Product Designer.
- Scope: research, information architecture, interaction design, MVP definition, and post-launch measurement strategy.
- Team: PM Stasha Markovska, engineering lead Dimitar Arnaudov, and co-research with Lisa Sawyer.
- Shipped scope: Connection Central v2 reached GA on December 17, 2024.
v1 had information, but not a map
Johnny’s v1 was not a failed product. It was a surface that had accumulated enough views that builders could see individual objects without reliably following their relationships. A connection, the resources it exposed, and the pipelines that depended on it could all be present while the answer to “what does this touch?” still required too much reconstruction.
That distinction mattered. I could have redrawn the existing tabs and called it cleanup. Instead, I treated the information architecture as the actual product decision: what should a person start from when they are trying to understand a complex, live integration estate?
Thirteen architects rearranged the problem
Before proposing a new structure, I ran a Structured Association study with 13 Solution Architects. They were close proxies for experienced builders who spent time in Pipelines and customer playback, not external customer participants. That limitation belongs next to the finding.
The pattern was consistent enough to change the top-level split. People described a resource — the connected system and what it exposed — before they described the internal object model around it. They wanted to understand what a connection affected, what depended on it, and where to inspect the consequences. The research did not validate every detail of the later UI. It gave me a credible reason to lead with resources.
This is a design specification of the resources side panel, not a claim that every illustrated screen shipped. The structure was the decision I protected: a builder should begin with the resource and keep its relationships visible while they inspect it.
Keep the map while inspecting one thing
The core interaction was a two-pane model. Navigation stayed on one side; detail appeared on the other. A builder could drill into a connection without losing the surrounding resource map. That was the part of the control-room metaphor that held up under real constraints: one view had to preserve context while the other did the detailed work.
The companion Pipelines view gave the domain a first-class list for answering a different question: what is running here? It belongs to the broader v2 design specification, with sample data plus Export CSV and 24-hour activity concepts deferred from the December GA release, rather than a labelled screenshot of shipped UI.
What survived the cut
Engineering had a fixed December GA date and limited UI capacity late in the work. The useful question was no longer which version I preferred. It was which decision would be expensive to reverse after launch.
I kept the resources-first IA and the two-pane shape. The accordion and popover detail did not survive. Neither did field-level detail, 24-hour activity emphasis, and CSV export. They were deferred, not quietly recast as shipped work.
That was a design decision as much as a delivery decision. A rich inline detail treatment demoed well, but it was not load-bearing for the first release. The two-pane model fit the available scope and preserved the context the research had actually made important.
Then we launched without a beta
On November 6, the PM and I decided to go straight to GA to meet the H2 date. There was no beta. This is the sort of sentence that looks decisive in a portfolio and feels considerably less elegant while making the call.
I wrote the v2 Pendo strategy so the team could read the surface after launch: a new-feature guide, onboarding walkthrough, feedback poll, and help update. Measurement was a follow-up mechanism, not a substitute for testing. The retained structure, the deferred details, and the no-beta decision are separate judgments. Blurring them together would make the launch sound more certain than it was.
The read-back measured reach, not relief
From July 20, 2025 through July 19, 2026, the reporting record shows 3,110 accounts using the Connection Central page. Separately, the measured entry-point population included 2,297 paying accounts, about 51.9% of the measured paying-account base. These are different populations. They describe reach through the measurement setup; they do not show revenue contribution, a redesign lift, or the quality of a builder’s decision.
I leave raw event volume out because automation contaminated that stream. I also leave the longer-window navigation funnel out of the headline because it does not establish a same-session usability result.
The unresolved question is the one that mattered when the project began: did the resources-first model stop people from making a mistake or help them recover when they were unsure? The data available here cannot answer that. I would still protect the launch date, but I would fight harder for a small moderated pass with the same Solution Architects before GA. The survey helped choose a model. It did not show where someone would hesitate, misread a status, or lose confidence in the middle of a real investigation.
Related work
Designing trust into Quickbase Pipelines covers the automation surface whose relationships Connection Central needed to make inspectable.
Designing Pave shows a related AI-builder navigation and recovery problem at public-launch scale.