PXP Protect
A React dashboard that turns a fraud model's probabilistic output into a decision an analyst can make in seconds — 5+ core KPIs, three Chart.js visualisations, 500k+ transactions at sub-50ms latency.
All figures and screenshots on this page are illustrative. No real transaction data is shown.
The problem
A fraud detection model is only as useful as the person reviewing its output. At PXP, the model produced a probabilistic judgment across hundreds of thousands of card transactions, but there was no purpose-built interface for reading it. The work was not “visualise the data” — it was: what does an analyst need to see first to decide whether to act?
Constraints
- Real-time data. The interface had to stay responsive at sub-50ms latency across 500,000+ transactions.
- A model outputs probability. A reviewer has to make a binary call.
- Internal tool, no existing design system to build on.
What I did
- Trained and tuned the fraud detection model, raising recall 20% over the legacy system while holding false positives under 1%.
- Designed and built 4+ dashboard pages in React, surfacing 5+ core fraud KPIs through three Chart.js visualisations.
- Built configuration controls so analysts could adjust the model's thresholds without going through engineering.
- Designed weekly and monthly summary views so patterns were visible over time, not just per transaction.
Outcome
Recall up 20% on the legacy baseline, with false positives held under 1%.