Case Study
Streamlining Validation for Data Scientists
Company
Cardlytics
Role
Sr UX Designer / Researcher
Type
Service Design, Desktop
Year
2020
The Brief
Cardlytics data scientists were manually hunting for sources of truth when validating transaction data — and waiting up to three weeks to find out if their work was right. The cost was measured in backlog items, wasted effort, and slow iteration cycles.
Choosing the Right Problem
Before designing anything, the team needed agreement on what to tackle first. During quarterly planning, I ran a prioritization exercise with design, product, and engineering — mapping every potential initiative against impact and effort. My job was to make sure there was a balance between user needs and business goals.
Into the Work
With priority set, I structured research in two phases: user interviews to understand attitudes and behaviors, and contextual interviews to observe data scientists on the job. The output was a current-state journey map built with all key stakeholders — creating shared agreement not just on what was broken, but on how it felt to experience it.
Drawing Clear Lines
The journey map surfaced overlapping pain points across multiple tools. Rather than guess at boundaries, I ran a card sorting activity with data scientists to let them define which needs belonged where. It gave us defensible product boundaries and eliminated the feature overlap debates that slow cross-team work.
Two Approaches
With boundaries clear, I moved into concepts. Two directions explored how data scientists might validate transaction strings — one surfacing algorithm matches with confidence scores for quick review, the other guiding them through a structured decision flow. We tested both to find out which fit how they actually work.
One Source of Truth
The final design gave data scientists a single place to search, review, and validate transaction strings — replacing the multi-system lookup that had been eating hours per day. Algorithm matches surfaced with confidence scores, recommendations guided the decision, and approved strings moved directly to the purchase graph without leaving the tool.