02 · AI · Visual exploration · Method

Singular: from zero to two decks in a week

How I used five generative AI tools in parallel to explore six visual directions and change the kind of conversation I was having with the stakeholder.

Drawing: a tower of stacked equipment and cables holds up a clean platform where a person watches a suspended cube.

00·

The brief

N5's CEO hands down a vision: Singular, a B2B customer engagement platform with AI for banks. There's technical architecture, there's a commercial hypothesis, there's strong intuition about where it sits in the market. What there isn't is a designed product, a visual narrative, or a concrete way to show the team and clients what the thing looks like.

The brief was clear: in one week, produce something that worked both to align internally and to sell. With no more context than that.

Opening slide: “A chatbot talks to the customer” crossed out and, next to it, “Singular talks to them on the bank’s behalf”. — The starting point: what Singular *isn't*. The proposition lands in a second.

01·

The first decision wasn't visual

The first thing I did wasn't open Figma. It was understanding the product. The first design decision I made changed everything that came after.

Singular isn't one product: it's two.

An operations panel for the bank's CRM Manager: high density, dark mode, real-time actions. And an executive panel for the C-level: read-only, top-down narrative, strategic synthesis. Treating them as one thing would have compromised both.

Separating them was the first deliverable of the week, and it was conceptual: I hadn't drawn anything yet. From there the work ran on two parallel tracks.

The first was the UX deck, the experience architecture of the whole system. How Singular works as a product, its three layers: the brain that decides, the voice that executes, the customer who lives the consequence.

Pencil illustration: a wired face, a figure speaking and a person looking at a phone, chained vertically and surrounded by annotations and technical diagrams. — The three layers of the system: the brain, the voice and the customer. Illustration generated with AI from the canonical prompt of the visual system.

Deck slide: a diagram of three chained boxes (Bank, North Star, Singular) and below them the three configurable layers, numbered. — "The bank defines the universe, the system activates the right version." On the right, that idea materialized in the conversation the end customer actually receives.

A hand holds a phone showing a messaging thread: the assistant proposes an alternative fixed-term deposit with its rate and the customer replies in two lines. — "The bank defines the universe, the system activates the right version." On the right, that idea materialized in the conversation the end customer actually receives.

02·

Mapping the space instead of defending one direction

The second track was the visual exploration of the admin. Instead of proposing one direction and defending it, I decided to map the whole space: five different generative AI tools, the same brief for all of them, six visual variants in parallel. Not to show I could produce a lot, but so the CEO could choose with real judgment instead of approving blind.

Every AI tool has a default voice that pulls toward a style. Fighting it wastes time; using it is what makes exploring six directions in a day possible.

One pulls toward clean fintech, another toward modern software product, another toward extreme density. Using each for what it does best turns a limitation into a method. For every variant I documented which tool was used, what prompt it got, what had to be fixed by hand and how fast it could be iterated. That documentation is part of the deliverable.

Variant	Visual direction	Tool
A	MOB Light — N5's own system, light mode	Claude Design
B	MOB Dark — N5's own system, dark mode	Claude Design
C	Liquid Glass — light physics, visionOS applied	Claude Code
D	Bloomberg Style — maximum data density	Lovable
E	Bugatti — fully monochrome, extreme discipline	Stitch
F	Lápiz Dark — editorial cream in dark mode	v0

Variant D: a data terminal with a sidebar of sections, small type and a full-width stacked area chart, with the percentages in a row at the bottom. — Variant D · **Bloomberg Style**. Maximum data density, no aesthetic concessions. The direction closest to the real banking context, and the one that ended up being chosen.

Variant A: a light dashboard with the health score set large, a line chart beside it and goal cards with a progress bar.

Variant B: the same dashboard on a dark background, with the goal cards turned into outlined blocks.

Variant C: the same dashboard with translucent surfaces and a saturated-color area chart. — Five of the six directions explored. Each one is a different hypothesis about what kind of product Singular is. All of them were produced in a day.

Variant F: a dark, restrained dashboard, in English, with the chart reduced to thin unfilled lines. — Five of the six directions explored. Each one is a different hypothesis about what kind of product Singular is. All of them were produced in a day.

7 days of work

5 AI tools

6 visual variants

2 decks delivered

03·

The outcome

Two actionable decks by the end of the week. The UX one to align the team on what is being built. The visual one so the CEO could make a real decision about aesthetics and direction, with full traceability of why each variant was considered and why it was dropped.

The methodological thesis behind all of it: AI as co-pilot, not as executor. Every design decision was mine; every asset produced went through at least one tool. The designer's role doesn't disappear. It becomes director, curator and critic of output.

The CEO walked into the final meeting with a real decision to make, and made it: the maximum data density direction. The closest to the real banking context, and the hardest to defend without having explored it against the other five.

What matters most to me about this case isn't the volume produced, but the kind of conversation it enabled. Instead of "here's what I decided", the dynamic became "here's the space, you decide".

And what it proved, at least to me, is that speed used well isn't a shortcut. It's what lets you explore the problem instead of solving only one version of it.

Outside the briefThe demo

Singular has a product in development. But when the CEO needs to give a keynote at Fintech Americas, slides aren't enough. The brief was to take a hard turn: forget the product for a moment and think like a show.

This demo isn't Singular — it's the concept car at the auto show. A working piece so the room can see where the system is heading before it exists.

Demo screen: a live map on the left, six agents in circles around a central one, and panels for goals, costs and human interventions on the sides. — The demo that accompanied the keynote. See live demo →

correction There was no brief, and at first I treated that as a problem. I lost two days waiting for the conversation to turn into a concrete request. It was never going to: when the brief comes from a vision and not from a requirement, the designer's job isn't to ask for clarity. It's to produce the options that make the vision decidable.

The first decision wasn't visual, and that's why this case exists. If I had started by drawing, I would have ended up defending my taste against the CEO's. Defining the exploration space first turned an argument about preferences into one about criteria, and that's the only kind you can document afterwards.

correction I didn't pick the tools out of curiosity. At first I used generative tools to produce the deliverable, and it came out fast but it came out looking like everything else. Here I used them to produce quantity: six directions, good and bad, in parallel, so I could choose with something in front of me. The decision stayed mine; what got faster was having something to choose from.

Ending up in a keynote isn't the outcome. The outcome is that the decision has traceability: you can reconstruct why that direction and not the other five. A blind approval would have ended up in the keynote too, but it couldn't have been defended when someone asked.