Making AI Health Insights Easier to Understand and Trust

Role

Product Design Intern

Timeline

Summer 2025

Team

2 Designers
2 Engineers
1 Product Manager

Skills

Product Design
User Research
Prototyping

Logging food was easy but understanding what it meant wasn't.

Flourish uses AI to connect what users eat with how they feel.

Logged meals plus symptoms, AI identifies potential patterns, personalized trigger insights

But those insights are only useful if users can log meals accurately and understand why the product surfaces certain patterns. My internship focused on strengthening both sides of that loop.

Every app I audited stopped at the data.

Across the 5 leading nutrition and meal-logging apps I audited, nutrition data was plentiful. What was missing was a clear connection between what I ate → how I felt → what I could do differently.

Competitive analysis quadrant chart — health data depth vs. next-step guidance, plotting Cronometer, ZOE, Function, NutriSense, and MyFitnessPal

More data wasn’t helping users make better decisions.

Across 15 interviews, the same pattern kept coming up.

"

With so many trigger foods and different levels of them, it's impossible to keep track on my own. I need technology to remember for me, but it's falling short at connecting that to my symptoms.

IBS patient

"

I log everything and still have no idea what caused my flare-up last week. It feels like I'm just collecting data for nothing.

GERD patient

Drowning in data

Dense nutrition breakdowns make "insights"
feel like a chore, not help.

Personal relevance mattered

Users wanted food triggers connected to their own symptoms, not generic nutrition rules.

No cause → effect link

Users couldn't connect what they ate to how they felt, making AI insights harder to trust.

Three ways to close that gap.

Explain the trigger

Help users understand why a food may affect them.

Make nutrition glanceable

Turn dense nutrition data into something users can read quickly.

Reduce logging friction

Let users correct meal details before bad inputs affect future insights.

I explored different ways to explain the trigger score.

Explaining the trigger score was the opportunity I spent longest on. Each direction answered part of the problem and left another one open. What shipped put the per-ingredient scores and the reason together.

Three lo-fi directions for the trigger summary card: severity labels giving each ingredient a rating, per-ingredient scores with no overall takeaway, and a meal-level score with an AI-generated reason but no per-ingredient attribution

The meal screen became the bridge from data to action.

Macro visualization

The visual cues were contradicting the data. With testing, I found that users interpreted the colors as judgments. 45g of carbs appeared red despite being well below the daily target, while 40g of protein appeared amber despite being close to its target. I explored single-color bars with visible targets to make comparisons clearer.

The same meal shown two ways: circular indicators, then horizontal bars with each macro anchored to its daily goal

Trigger Score

The card had to carry the whole story without burying it. Collapsed, it leads with the score and the reason behind it, so the takeaway is immediate. Expanded, each ingredient opens into its own macros and score, so users can see exactly what pushed the number up and what to swap.

The trigger summary card collapsed, leading with the score and its reason, and expanded, showing each ingredient's macros and score

Edit Logged Meal

Logging mistakes, like the wrong meal type or time, would carry into future trigger insights. Tapping the meal name or its chip opens a quick edit, so users can fix details in a few taps instead of re-logging the meal.

All together

Demo video coming soon

Designing for the edge cases I missed.

During engineering handoff, a developer flagged that longer meal names were breaking the layout. I revisited the component in Figma, allowing names to wrap to two lines while keeping the edit icon and timestamp aligned. I documented the behavior for implementation so users could still recognize and verify their logged meals.

It shipped, and interpretation time dropped.

The redesigned meal experience shipped in Flourish AI and was featured in the product's App Store preview. In a comparative usability study with 8 beta users, the redesign helped users interpret meal insights faster and more accurately identify their personal food triggers.

47% reduction in time spent interpreting meal insights, 34% increase in users correctly identifying personal food triggers, shown against the redesigned meal detail and trigger insight screens

I got more comfortable being wrong.

I went into this project confident in some of my design decisions, only to find that users interpreted them completely differently. I learned to focus less on defending my ideas and more on understanding why something wasn’t working. I am so grateful to be part of building an app that directly helps people with problems that many struggle with.

Next Up…

Carbon Sustain

Making Carbon Data Actionable

Amazon Music
Amazon Music

Consumer & Artist Collaboration