Why Tinder's CPO Starts Every AI Prototype With JSON
Ravi Mehta
CPO Tinder; founder Outpace; AI prototyping expert
SEP 29 2025
The Method
JSON First, UI Second: The AI Prototyping Rule
"Before I draw a single wireframe, I write out the data model in JSON. If I can't describe the AI behavior in structured data, I don't understand it well enough."
JSON spec forces clarity: what data does the AI consume? What does it output?
Design comes after logic — not before — in AI products
The prototype question: does the AI make the right decision on this JSON input?
Ravi's insight: most AI product failures are data model failures dressed up as UX failures
Step 2: Validate the AI logic with real inputs before touching the UI
Step 3: Design the UI as the output layer, not the starting point
Step 4: Evaluate against JSON spec before user testing
The Tinder lessonTinder's matching algorithm is the product. The swipe UI is the expression of that algorithm. AI products work the same way.
Prototyping AI Products
What Changes in AI Product Design
Old prototype: Wireframe → feedback → redesign → build
New prototype: JSON spec → AI logic test → UI design → user test
The shift: The earliest prototype is a spreadsheet or JSON file, not a Figma screen
The evaluation: Does the AI decision match what the user would expect?
The edge case test
Write 20 edge cases in your JSON spec before you build. If the AI fails on 30% of them, the product will fail too.
The explanation design
AI products often need to explain their decisions. Design the explanation before designing the action.
Playbook
Prototype AI Products Better
Always start with the data model: what does the AI need to know to make a good decision?
Test AI logic before testing UI: is the decision right? Is the explanation clear?
Build evaluation into the prototype: 20 test cases before you show it to users
Design for the failure case: what does the UI show when the AI gets it wrong?
The CPO perspectiveRavi now coaches product leaders at Outpace. His #1 AI product feedback: teams prototype the UI first and validate the AI logic last. That's backwards.
Contrarian
AI Product Design Myths
✗Design the UI firstINSTEAD →✓ Design the data model first. The UI is the wrapper for the AI decision — design the decision first.
✗User testing validates AI productsINSTEAD →✓ User testing validates UX. JSON spec testing validates AI logic. Both are required.
✗AI products should hide their logicINSTEAD →✓ AI products should be transparent about their reasoning. Explainability is a UX feature.
✗Iterate the model when users complainINSTEAD →✓ Iterate the data model when users complain. Most AI UX problems are data model problems.