2016–2018

Feedzai

AI/ML Risk Management and Fraud Prevention. Two flagship products, a $50M valuation milestone, and the start of a long interest in designing for AI-powered systems.

Fintech Risk Intelligence AI / ML B2B SaaS
Abstract diagram: a timer and code editor feeding into a browser buy flow, connecting through to a global alert siren
Role Lead Product Designer
Company Feedzai · Lisbon, PT
Period 2016–2018
Stage Scale-up · $50M valuation
Focus Case Manager · Insights · UX Research · Design Team
The Situation

Sophisticated technology, a legacy product, a team of two

Feedzai builds AI-powered fraud and risk detection for banks and payment companies. The technology was sophisticated. The product analysts used to act on it, a legacy Alert Manager, wasn't. It didn't scale, and it couldn't handle where the business was heading.

I joined as the second designer, tasked with building the design and research practice from scratch. Analysts were making hundreds of judgment calls a day, each with real financial and reputational consequences.

My Remit

Design the product. Establish the practice.

  • Design Case Manager from scratch: risk and case management with rich data visualisation and self-service configuration
  • Grow the design team from two people to eight
  • Build a UX research practice: workshops, personas, recurrent usability testing
  • Make ML output legible and actionable without a data science background
Defining the Problem

Structure before screens

Case Manager needed a structure analysts could hold in their head: nine sections, from Search and Overview to Rules and Admin, mapped before a single screen was drawn. Paper sketches and wireframes tested that structure against real flows, account recovery, rule configuration, audit trails, before anything reached engineering.

Case Manager information architecture: Search, Overview, Alerts, Entities, Analytics, Rules, Pos/Neg, Admin, Profile Paper sketches and user flows, followed by mid-fidelity wireframes of the Self-Service Rules and Case Manager Helper screens Transaction Statistics report and Business Overview dashboard: alert volumes, approval rates and team performance
The Work

From legacy alert queue to decision workspace

With the structure validated, high-fidelity screens replaced the legacy Alert Manager one workflow at a time, each tested through moderated sessions, SUS scores and attrakdiff, so we had signal throughout the build, not just at the end.

The interface followed the analyst's decision process, not the data model. Insights translated ML output into a form compliance could read and trust: specific enough to be meaningful, honest enough not to overstate certainty.

Alert Details screen: customer profile, payment methods and activity log, alongside the Alerts queue with score, status and entity columns Gross Fraud Performance dashboard: totals by approval type, trend over time, and fraud by region on a world map

Case Manager: evaluative research

Problem. After the formative work, we needed to evaluate usability, not assume it.

Solution. Recurring usability tests with a fixed internal cohort: moderated interviews, task success metrics, SUS and attrakdiff scores.

Usability testing sessions (July 2016 + January 2017)

SUS scores of 91.5 (A, Excellent) and 88.5 (B, Excellent), and Attrakdiff pragmatic quality, hedonic quality, attractiveness and goodness ratings
Key Decisions

Three calls that shaped the outcome

01

Design the AI as a collaborator, not an authority

The model score was visible, never the headline. Analysts saw the evidence first, the model's read second. They caught model errors a more automated interface would have missed.

02

Audit trail as a design constraint, not an afterthought

Every decision here had regulatory weight. Building the audit trail in from day one, not bolting it on as a report, shaped how every interaction was structured.

03

Research as a product input, not a UX deliverable

The 40+ interviews weren't a slide in a deck. They fed straight into prioritisation, IA, and interaction patterns, so the product matched how analysts actually worked, not how we assumed they did.

Components, documentation and academy

Scaling the team meant documenting the system too. A UI kit, and a Feedzai Academy training platform built with the technical writing team, kept the design team and our customers working from the same reference.

Feedzai UI kit library and documentation, the Feedzai Developers portal, and Feedzai Academy: a training platform for onboarding co-workers and customers
Outcome

Products that analysts trusted

In a domain where bad decisions carry real financial consequences, trust was the only measure that mattered. Case Manager and Insights became defining products in Feedzai's commercial expansion.

$50M Valuation milestone at time of launch
2→8 Design team growth during tenure
40+ Research sessions with fraud analysts
2 Flagship products shipped end to end
SUS 91.5–88.5 Excellent range across testing rounds
Reflection

What this changed

Feedzai is where I first had to design alongside machine learning, not just in front of it. Making opaque model output legible without eroding human judgment is a problem I've carried into every AI-adjacent project since, and it's more relevant now than it was then.

The model score was never the headline. Evidence first, then the model's assessment.