CAP Internal use · being validated · screens with blurred data
Cap Intelligence — Turning market data into traceable decisions
An internal portal, started on my own initiative, for Liderança's product team to read the capitalization market with sources, caveats, and hypotheses in plain sight.
Business and Product have already used the portal to decide to investigate two hypotheses
CAPThe portal
From the number to the decision, with what could weaken it in plain sight.
Screens of the internal portal (Nov 2026). Values, hypotheses, and internal readings were blurred; the market data is SUSEP's public database. The interface is in Portuguese.
Context
Liderança Capitalização had a lot of data — the public SUSEP database, inflation from the Central Bank, behavior on the website and the app — but few people could interpret it and turn it into product decisions. I started Cap Intelligence on my own initiative, for the product team.
Problem
Help the team decide where to invest — where the market grows, where Liderança is absent — without turning a possible reading into a fact. Public data has traps: reversals, changes in classification, partial coverage. A number without context leads to a wrong decision that looks like evidence.
My role
Product Designer Lead & AI Builder, on my own initiative: diagnosis, PRD, information architecture, interface, and implementation with AI agents in Cursor. Decisions that are not design decisions — data authorization, AI use, hosting — stay with Business, Finance, IT, and Compliance, and the portal records what each one is blocking.
Decisions
Every number with its source and data version. Deterministic, tested calculations, without AI. Caveats generated by fixed rules, next to the analysis they weaken. Separate what we know, what we believe, and what we do not know. Hypotheses and decisions recorded and re-evaluated on every data update. An AI agent that reads only the public report; for internal data, it only translates the question and the portal calculates the answer. Navigation by the team's questions, not by the sources.
Solution & artifacts
Screens and project details.
From data to decision, without hiding what the number does not say
The portal does not try to guess the market. It shows where the market moves, says why the reading may be wrong, and records what the team decided to do with it.
- Traceable — source, period, and data version on every number.
- Honest — a hypothesis shows up as a hypothesis, with what is missing to validate it.
- AI with a boundary — the agent reads the public report; internal data does not leave the portal.
Data, reading, decision
-
Data
- SUSEP and the Central Bank
- a version for every collection
- checked against the source
Decision: every number with its provenance
-
Reading
- fixed rules, no AI
- we know / we believe / we do not know
- caveats next to the analysis
Decision: what is measured kept apart from what is inferred
-
Decision
- recorded hypotheses
- re-evaluated on every collection
- a decision with its rationale
Decision: a decision you can trace
What we discarded
Problem
Help the team decide where to invest — where the market grows, where Liderança is absent — without turning a possible reading into a fact. Public data has traps: reversals, changes in classification, partial coverage. A number without context leads to a wrong decision that looks like evidence.
Result so far
Two hypotheses taken to a decision by Business and Product on Oct 29, 2026, with the rationale recorded. The portal is still being validated.
Tensions
- Speed × rigor
- AI × traceability
- Insight × hypothesis
- Internal data × AI provider
Impact
On Oct 29, 2026, Business and Product used the portal to decide to investigate two hypotheses, with the rationale recorded — and that is the metric the product pursues: hypotheses decided, not screens or charts. The portal still runs on the internal network, without authentication, and most screens have not been tested with users; the human review of caveats depends on authentication.
What I learned
Having data is not having information. The design work was to separate what is measured from what is inferred — and to show what the number does not say before someone takes it into a meeting.