Personal Finance Management

A research-led personal-finance app for Caixa Econômica Federal’s customers, centred on the one job the evidence kept surfacing: helping people save. I took it end-to-end during the 2020 lockdown: discovery, research synthesis, IA, a visual system, a hi-fi prototype, and two rounds of usability testing.

Caixa was not a client. I picked it because pandemic-era load made the app’s failure modes visible to every Brazilian at once. The brief inverted the usual move: instead of stacking features onto the existing app, build a smaller one alongside it. Let the legacy app keep the transactions. The new surface centres on saving toward goals (called Sonhos in the product), spending guidance, and short financial-literacy content.

The new app · goal-saving home, spending insights, financial assistant, cashback · hi-fi prototype in Brazilian Portuguese

A bank people already hate to use

In 2020 the federal government routed pandemic emergency aid (auxílio emergencial) to tens of millions of unemployed and informal workers, and Caixa was the bank chosen to pay it. Overnight, people who had never touched a banking app were forced into one to reach money they urgently needed. The app buckled: crashes, failed logins, screens nobody could navigate, and a flood of one-star reviews on top of an app that already crashed on a balance check. The agency queues made the news every week. The App Store reviews became a public ledger of failure. That setup was the brief.

Caixa’s app rated 2.x stars on both stores. I pulled both review corpora (App Store and Google Play, August 2020), tagged each entry by theme, and ran a frequency analysis. Both stores said the same thing: the app doesn’t work, and users don’t trust it to do the basics.

Review wordcloud · Caixa app · App Store + Play, Aug 2020 · Dominant complaint terms picked the direction

Strip the neutral nouns and the most-frequent phrases read like a list of usability defects: "temporarily unavailable" ("temporariamente indisponível"), "system unavailable" ("sistema indisponível"), "doesn’t work" ("não funciona"), "I can’t" ("não consigo"), "won’t open" ("não abre").

Why this case

Two more reasons it was worth picking, beyond the public failure.

Sole designer, every stage

A self-directed bootcamp capstone (UX Unicórnio 1.0). I owned every stage, from problem framing through to usability testing. No client, no team. The bootcamp critique loop was the only sounding board. Useful, as it happened: every decision had to defend itself without an account manager or PM to absorb it.

Evidence from four streams

No client meant no budget, no participant panel, no incentive pool. Research came from what I could reach: public review corpora, two interviews with active Caixa account holders, a short survey of acquaintances, and a belief-mapping exercise I re-sorted each time new evidence came in.

One persona, deliberately not the heaviest pain

I started with two protopersonas. After the interviews, the second one collapsed into the first. Same goals, same blockers, just less specific. Júlia stayed. She’s 29, single, works in advertising, makes R$3,000 a month. She is solvent, digital-native, and gets through the month fine without much help. Picking her over a higher-pain persona raised the bar on the brief: there was no "broken" to fix, only friction to remove and behaviour to nudge.

Persona Júlia · Digital-native, solvent, wants to plan better · A solvent, digital-native persona chosen for friction

From "I should save" to "I did"

A six-stage journey from intent to goal-hit, mapping what Júlia does, thinks, and feels at each step, and where the design can intervene. The opportunities row is the synthesis: the moves I carried into the build.

Scroll the map sideways →

Customer journey · Júlia, intent to goal-hit · The opportunities row is where the design earns its place

What made the MVP

An impact × effort 2×2 to keep the scope honest. Anything in the high-impact / low-effort quadrant made the MVP. Everything else was either parked, cut, or named explicitly as a v2 idea so I wouldn’t accidentally build the whole thing around it.

Impact × Effort matrix · The high-impact / low-effort quadrant became the MVP · NFC-cashback parked to v2

Paper first, on purpose

Roughly 30 screens on paper before Figma opened. The home layout went through four paper versions I’d otherwise have built in Figma and grown attached to. Paper is faster to throw away, which is why it earns its keep at this stage.

Paper sketches · ~30 screens · Navigation and key flows resolved before pixels

Lo-fi prototype · Greyscale Figma frames · Structure validated before colour

Brand orange, claimed properly

The visual system pulled the primary colour from Caixa’s existing brand orange. Two reasons: it keeps the work recognisably Caixa, and the legacy app uses orange sparingly and badly. Claiming it as the CTA colour made the brand feel intentional instead of incidental.

An earlier iteration leaned on a blue/orange split closer to Caixa’s full logo. Dropping the blue cleaned the system up and made the CTA orange unambiguous. There is one "act" colour now, and it is always the same one.

Colour

Three primaries (brand orange plus a lighter and darker variant), a five-step neutral ramp, and exactly two feedback colours, one error and one success. Anything more would have been decoration.

Typography

Roboto, picked for neutrality. Five heading levels, two paragraph weights, one small caption. The smallest scale that could hold the IA without compromise. H1 sits on the home Sonhos header. Everything underneath tunes for scanability.

Buttons

Three roles, four states each. CTA stays reserved for orange: the moment the user is supposed to act. Secondary handles "not yet" branches. Primary handles inline confirms inside flows. Disabled never relies on colour alone; the contrast is structural too.

Sonhos replaces transactions as the home

The hi-fi prototype centres on goal-saving. The home tab is goals (Sonhos), not payments (Pagamento). Each goal is a CDB-backed savings target with a progress bar, an estimated yield, and the next-best action surfaced inline. The other tabs are second-class by intent. They support the saving job rather than compete with it.

Interface in Brazilian Portuguese; captions translated.

Sonhos home · Goals replace transactions on the home tab

Sonho detail · Editable target, progress, and next-best action inline

Assistente Financeiro · AI builds per-category saving targets from past 6 months of spend

Cashback · Earned at partner stores, routed into a chosen Sonho

Pagar · Present, not promoted · Legacy app stays the system of record

Blog Financeiro · Bite-sized literacy content · Reading-time tags lower the entry cost

What broke, what changed

Two rounds, more than six participants between them. Verbatim notes for every session. I paired each finding with a concrete design change; findings that didn’t lead to one stayed on the open-questions list instead of quietly disappearing.

Written in 2020, still holds

After shipping the prototype I logged a self-critique while the project was still fresh. Rereading it years later, almost every item is still right. They’re listed below as I wrote them, lightly translated.

Reframe the project objective as a problem statement

"50% of users save 30% in three months" was a wish, not a measurable outcome. A fair target for a prototype is whether participants understand each screen and would open it twice. I’d cut the wish-target on day one.

Drop the temporal axis from the journey

My journey is structured chronologically by app stage. It should be structured by experience: what the user is trying to accomplish at each step, regardless of where in the product they are. Map experience, not app surface.

Accessibility

I validated the dark theme and hot orange against brand match, not against WCAG contrast at every body-text size. Touch targets, focus states, screen-reader labels and motion-reduce behaviour didn’t make it into the design-system spec. A few surfaces (the cashback partner cards, the meta-text on Sonho rows) almost certainly fail AA on dark backgrounds.

There was also no support for text enhancement: the design assumed one fixed type size, with no honouring of the OS text-size setting, no Dynamic Type, and no sane minimum body size. Anyone who scaled their system font up (exactly the older or low-vision users a public bank has to serve) would have broken the layout. Today that’s table stakes, and something I’d build in from the first screen.

A production version would treat accessibility as a token-level requirement, not a polish step. Contrast pairs would live inside the colour token. Focus rings would be a primitive. Type would scale with the user’s system setting and reflow without clipping or truncation. Screen-reader semantics would get specified per component, before any screen is composed.

Push financial education earlier

The Blog Financeiro tab arrived late in the build and stayed at sketch-level depth. The research signal was clear that financial literacy mattered as much as tracking. I’d treat it as a peer feature, not a side panel.

What still applies

This was the first project I owned end-to-end. The research depth was disproportionate to the deliverable: six weeks of discovery against four weeks of design. For a concept where the brand itself was the biggest unknown, that ratio was the right call.

Real complaints beat invented personas

The store-review wordcloud did more strategic work in a day than any persona I wrote that month. When public evidence exists, mine it first. It forces the problem statement to defend itself against real language instead of synthesised pain.

Bias is a designed surface

"I already hate it" ("Já odeio") was design feedback. No screen-level fix could undo the brand-bias finding. Putting it in the case rather than burying it is part of the work.