User Data Management System

Learning how documents, data, and AI extraction worked

The first stage focused on deeply understanding the data model and document structure behind the product. This phase was essential to ensure that the UI accurately reflected the system logic and supported all required scenarios.

• Analyzed the tax documents commonly used by the system to understand what information was extracted.
• Reviewed back-end JSON files and schemas to map all available fields and their relationships.
• Identified how personal data, household information, and dependents were structured within a tax declaration.
• Collaborated with engineering to understand how AI extraction worked, including confidence levels and known limitations.

Reviewed tax documents to understand the information extracted by the system.

Collaborated with engineering on AI extraction, confidence levels, and known limitations.

Identified how personal data, household information, and dependents were structured.

Analyzed back-end JSON files and schemas to map available fields and relationships.

Translating system complexity into user needs

Once the data structure was clear, I focused on defining the most important user scenarios, especially those involving uncertainty or errors. This included enabling users to edit incorrect AI-extracted information, understand which document and section each data point came from, and manually add missing information when needed. A key requirement was clearly differentiating AI-generated data from user-entered data. I also considered complex household setups, including multiple dependents, ensuring the experience scaled across different tax situations.

By understanding different information hierarchies, such as family members and income sources, I organized categories in a logical way, essentially creating an information architecture based on the underlying data.

I held multiple meetings with the back-end teams to understand how the data was structured, stored, and grouped in different ways. The goal was to design an interface structure flexible enough to adapt to multiple use cases.

Exploring solutions quickly with the team

To explore and validate ideas efficiently, I created multiple design proposals using Figma and V0. These prototypes allowed me to compare different ways of displaying AI-extracted data, test visual indicators for data origin, and validate information hierarchy and editing flows. Designs were reviewed collaboratively with product and engineering during working sessions, helping the team converge quickly on a consolidated direction.

I used v0 as a tool to quickly prototype ideas and validate concepts with the team.

The designs were later moved to Figma to validate and apply the design system components.

Designing for rare but high-impact failure cases

A critical phase of the project was designing for failure and low-probability edge cases. Although complex extraction errors occurred in less than 0.5% of cases, the product needed to remain fully usable even when AI could not be trusted. I designed fallback flows that allowed users to complete the entire experience manually, along with clear messaging when extracted data could not be validated. Special attention was given to safe progression states to prevent data loss in complex error scenarios, ensuring the experience remained trustworthy even in worst-case situations.

Users can see when the AI detects multiple possible values and are prompted to verify and select the correct one.

Scaling the solution with the design system

After validating the core flows, I moved into full design execution. All screens were designed using the existing design system, which I extended where necessary to better communicate field states, data confidence, and the distinction between manual and AI-generated inputs. I also built a functional, high-fidelity prototype to support usability testing. These updates helped improve both the clarity of the experience and the scalability of the design system.