Designing an AI-powered Recommendation System that turns health reports into personalized action
Diagnostics • consumer app
00
Context
00
After successfully helping users understand their lab reports through health insights, we noticed something interesting.
Users appreciated knowing what was wrong, but they still struggled with the most important question:
"What should I do now?"
Reading insights doesn't improve health. Taking action does.
This realization led us to design the second phase of Health Insights Hub—a recommendation system capable of converting medical insights into personalized health plans.
Unlike traditional recommendation engines built using rigid medical rules, our goal was to build a scalable AI-assisted recommendation framework that could continuously evolve with new medical knowledge, user preferences, and healthcare services.
01
The Problem
01
Initially, recommendations were generated through a rule engine.While medically correct, the experience had several limitations. Recommendations looked generic.
Every user received almost identical advice. There was no explanation of why a recommendation appeared.
Adding new recommendation pathways required engineering work and medical validation every single time.
Most importantly, recommendations felt disconnected from the user's personal health story.
Interview: 10-15 mins
35 Calls
Interview: 30 min
3 in person
The objective was to understand how people perceived recommendations after understanding their reports.
The findings completely changed our direction.
Supplements were viewed as medical interventions rather than informational suggestions.
This meant recommendations required a stronger trust layer.
Doctor validation became equally important as AI intelligence.
Users preferred:
- visual guidance
- structured plans
- simple checklistslear priorities
Research showed recommendations needed better information architecture, not more information.
This suggested that behavior change should become the first recommendation layer before supplements or medications.
Users repeatedly asked:
"Why am I seeing this recommendation?"
We realized personalization wasn't just about generating different advice.
It was about making users understand why that advice belonged to them.
02
Defining the Design Challenge
02
We weren't designing a recommendation page. We were designing a recommendation platform — one that had to satisfy three competing constraints at once.
Every design decision lived at the intersection of medical accuracy, personal relevance, and product scalability. Lean too far toward medical correctness and recommendations felt generic. Over-index on personalization and you risk unsafe advice. Optimize for product reuse and you lose the human feel.
03
Existing Approach
03
The conventional healthcare approach looked like this:
Medical team creates protocols
Engineering builds rule engine
Design creates screens
Recommendation appears
Every new recommendation category required rebuilding logic. This process took months. It also limited experimentation.
Reframing the Problem
Instead of asking:
"How do we generate recommendations?"
we asked:
"What are recommendations fundamentally made of?"
04
Breaking Recommendations into Building Blocks
04
We analyzed hundreds of recommendation examples across medicine, nutrition, preventive care, and consumer health.
Instead of treating recommendations as paragraphs of text, we decomposed them into reusable UI patterns.
Each pattern became a reusable design component. We called these our recommendation "lego blocks."
05
Teaching AI the Design System
05
Rather than asking AI to invent interfaces, we taught it how each component worked.
Every block contained metadata describing
Purpose, interaction, medical intent, constraints, supported content, hierarchy, when it should appear, when it should never appear
Instead of generating UI,
AI selected and assembled existing product components.
This dramatically improved consistency while keeping recommendations personalized.
What we shipped in version 1:
All components rendered and choosen by AI based on the lab reportpowered by AI
06
Designing for Trust
06
The biggest product challenge wasn't personalization. It was trust.
Research consistently showed users trusted doctors significantly more than AI for higher-risk decisions. We redesigned recommendation hierarchy around trust.
Users are more comfortable acting on low-risk suggestions. The system surfaces those first and gates higher-risk actions behind doctor consultation.
Design implication
The AI doesn't just select what to recommend — it determines how to present it. Diet cards are actionable with one tap. Supplement cards always pair with a "Consult your doctor" prompt. Clinical suggestions link to the biomarker detail page, never to a purchase flow.
07
Key Design Contributions
07
Led end-to-end product design for AI-powered recommendations.
Conducted synthesis of qualitative user research to identify trust and personalization gaps.
Defined a modular recommendation design system reusable across multiple health surfaces.
Partnered with AI engineers to establish a component-driven prompting strategy instead of free-form generation.
Designed contextual recommendation experiences embedded directly within biomarker detail pages.
Created scalable interaction patterns that balanced AI flexibility with medical safety.