Marcel M. | Product design portfolio

// AI-workflow

Understanding where AI creates value in the design process.

timeline

2026

Domain

AI workflows

Scope

Research, experimental

outcome

Prototypes

Context

Since finishing my last assignment, I’ve been exploring how AI is changing the way digital products are designed and built. Most AI conversations I see focus on generating screens or writing code.

I wanted to know where AI removes friction in the product development process without compromising quality.

// Goal

I wanted to see how AI could help product teams spot risks, uncertainty and opportunities across research data, journey maps and service documentation.

// Scope

I treated AI as a product problem, not a design tool. I wanted to understand where it could support research, discovery, analysis and implementation, the workflows that eat the most time on a project.

// responsibilities

AI engineering.

Problem

Across projects at the Dutch Tax Authority and DigiD, I kept hitting the same problem. Research findings, journey maps and service documentation lived in separate places. To see where users struggled and what evidence backed each conclusion, I had to manually connect information across every source.

Opportunity

An AI assistant could analyse research notes, journey maps and service context together, then surface risks, opportunities and open questions.

FIG_01 :: Overview of the workflow

Hypothesis

I believed structuring existing research would cut the effort needed to analyse complex service journeys.

// SUCCESS CRITERIA

Traceability

You can trace every insight back to its source: the original quotes, observations and journey steps.

Research synthesis

Teams cluster findings, spot patterns and generate opportunities faster.

Evidence-backed

Every opportunity stays linked to the research it came from. Teams can validate a recommendation against real user evidence.

Prototype

To test the concept, I built a lightweight prototype using Python, the Claude API and structured markdown files. The goal was to validate whether AI could connect journey steps and research findings into a single analysis.

FIG_02 :: A lightweight Python application powers the workflow.

// under the hood

High-level concept

How the application operates,

Technical workflow

It takes raw research and outputs risks, opportunities and open questions.

Evaluate reliability

Anyone can generate insights. Trusting them is the hard part. I tested the prototype against incomplete and vague research data to see how it responded.

FIG_01 :: Overview of the workflow

FIG_02 :: Incomplete inputs tested the assistant's ability to identify evidence gaps.

FIG_03 :: Incomplete inputs tested the assistant's ability to identify evidence gaps.

FIG_04 :: The assistant reduced confidence scores and generated follow-up questions when evidence was insufficient.

Reflection

I came away thinking differently about AI. The real challenge is designing systems that help people understand where insights come from and when to trust them.

// key learning

Give people a way to check the AI's reasoning before they trust its output.

// What worked

✓ Structured outputs made findings easier for review.
✓ Open questions helped identify gaps requiring more research.
✓ Linking insights to supporting evidence improved transparency.

// What didn't work

- Incomplete research data reduced output quality.
- Weak evidence occasionally led to overly confident conclusions.

// OTHER PROJECTS