All projects
ITAÚ / AI PRODUCT CONCEPT · 2026
i.agora
From diagnosis to a next step.
I designed the journey, AI interaction and visual system for a financial experience that turns a diagnosis into a plan people can review. I refined the prototype individually after the hackathon.
Hackathon concept · synthetic data · later refinement. Not launched by Itaú.
Try the prototype
Context before advice
Start inside the existing banking conversation, without creating another app.
A plan people can review
Show concrete amounts and let the person accept or adjust the proposal.
Projection, not a promise
Separate deterministic calculations, AI interpretation and human choice. Proposed architecture, not validated in production.
< 01 > CONCEPT
Make the future visible. Preserve choice.
During Itaú’s AI agent hackathon, we explored how transaction context, deterministic calculations and generative AI could help people facing financial vulnerability move from diagnosis to action.
Maria is a fictional character. Screens show the prototype refined after the hackathon; all financial amounts are simulated.
< 02 > PROBLEM
The balance was a snapshot. The trajectory was missing.
What we found
We started with the hypothesis of helping people in debt. Analysis of the synthetic dataset revealed a broader pattern: ending the period close to zero or overdrawn, even without missed payments.
What this changed
The focus shifted from reacting to debt to making financial consequences understandable before the next shortfall. Understand the context, anticipate scenarios and support a choice.
Reactive banking → preventive financial guidance.
How to interpret the 30–34%
This range reflects only the team’s directional interpretation of the hackathon dataset. It is not an estimate of Itaú’s customer population, an official statistic or evidence of statistical significance. The signal helped frame the problem; it does not demonstrate the concept’s effectiveness.
SYNTHETIC DATASET
~1
k
Customers in the challenge sample
Hackathon sample · not Itaú’s customer base
SYNTHETIC DATASET
~98
k
Transaction records analyzed
Data provided for the challenge
DIRECTIONAL SIGNAL
30–34
%
Of the sample showed signs of vulnerability
Team interpretation · not an official Itaú statistic
< 03 > PRODUCT DECISION
One person. One problem. One plan.
The initial idea combined multiple futures, agents, goals, gamification, products and rewards. The challenge’s time limit called for a story people could understand and try.
INITIAL EXPLORATION
Broad ecosystem
Multiple futures, agents, goals and rewards.
REFINED DIRECTION
One next decision
Context → adjustable plan → follow-up.
Later refinement reduced distractions and kept Maria’s story in focus.
< 04 > JOURNEY
From context to choice. From choice to follow-up.
Reality → conversation → diagnosis → plan → commitment → follow-up.
DECISION 01
Start with the conversation that already exists.
Maria arrives through i.ai, within the context of her account. i.agora appears as a specialist capability, avoiding another app or another destination to learn.
DECISION 02
Replace “spend less” with a proposal people can adjust.
The plan presents concrete limits and amounts. The 50 / 30 / 20 reference is adapted to the character’s situation; Maria chooses to accept or adjust it. It is a simulation, not a universal formula.
DECISION 03
Show what is a commitment and what is still a projection.
The optional card appears after acceptance and does not reveal financial amounts. Follow-up distinguishes the projected R$620 reserve from January’s progress, which is still unconfirmed. No transfer is automatic.
< 05 > AI AND TRUST
Calculate with rules. Explain with context.
In the proposed architecture, transaction data provides financial context. Deterministic calculations and simulations produce the numbers; the LLM supports interpretation, conversation and explanation. The interface turns them into a proposal the person can review.
01
Data
Transaction history
02
Context
Income and commitments
03
Calculations
Rules and simulation
04
Conversation
Interpretation with an LLM
05
Interface
Informed choice
AI explains possibilities. The person stays in control.
Autonomy
Accept or adjust the plan. Never decide or move money on the person’s behalf.
Transparency
Make assumptions, simulated data and projections explicit. Do not promise outcomes.
Care
Avoid guilt, fear or shame as persuasion. Provide a path to human support.
My contribution and the limits of delivery
My work connected problem framing, the journey, AI interaction, prototyping, the visual system and proposed metrics. I created Figma components and tokens based on the existing app experience, explored prompts and conversational behavior, and independently refined the prototype after the hackathon.
The team included Engineering, Data, Product and Design. Data analysis and modeling, backend, the complete architecture and GCP infrastructure belonged to other disciplines. Separating calculations from language is a principle of the concept; it was not validated in production.
During design, AI helped explore alternatives, structure prompts, prototype and iterate on the front end. Scope and experience-quality decisions remained under human judgment.
< 06 > PROPOSED VALIDATION
A functional experience. Hypotheses still open.
The concept did not enter production. There are no measured customer or business outcomes. The metrics below were proposed to test whether the experience helps people understand, decide and follow a plan.
HYPOTHESIS / QUALITY
>80
%
Success in relevance, clarity and context
Target proposed in the challenge · not measured
Action after simulation
Can the person turn the scenario into a next action?
Adherence to the plan and goals
Does the agreed plan translate into progress over time?
Relevance and conversion
Products enter the evaluation only when they make sense for the context and journey.
Business opportunity follows value for the person.
< 07 > RETROSPECTIVE
Less would have made the idea stronger.
What worked
The problem emerged from the challenge data. Collaboration across disciplines and rapid prototyping brought an abstract idea closer to a functional experience.
What I would do differently
I would choose a central hypothesis sooner, define success criteria before the final phase and cut parallel features to protect the clarity of the presentation.
The post-hackathon refinement followed that lesson: reducing distractions, strengthening Maria’s story and clarifying the proposal and its limits. These screens are from that later iteration, not the exact version evaluated in the competition.
Make complexity clear enough to act on.
Try the prototype







