Styleum

The Problem

Existing wardrobe apps are broken. They demand 20+ minutes of tedious photo uploads before you can even get started. Most users abandon before seeing any value. And the AI recommendations? Either too expensive to scale ($0.02+ per call) or too generic to be useful.

The Insight

Fashion is deeply personal, but the technology to deliver personalized styling doesn't need to be expensive. By routing simple requests through lightweight models and only escalating complex style decisions to GPT-4, I cut costs by 10x while maintaining quality.

The Solution

Hybrid AI Pipeline

The key innovation is a two-stage AI pipeline. Stage 1 uses a fast, cheap model to classify the request type and extract key parameters. Stage 2 routes only the complex requests to GPT-4 with rich context.

async function routeStyleRequest(request: StyleRequest) {
  // Stage 1: Fast classification (< 100ms, $0.0001)
  const classification = await classifyRequest(request);

  if (classification.complexity === 'simple') {
    // Handle with rules + lightweight model
    return generateSimpleRecommendation(request);
  }

  // Stage 2: Complex requests get full AI treatment
  return generateAIRecommendation(request, {
    model: 'gpt-4-turbo',
    context: buildRichContext(request.userProfile),
  });
}

3-Minute Onboarding

Instead of demanding a full wardrobe photo shoot, Styleum starts with style preferences and body measurements. The AI builds your profile iteratively—every interaction makes recommendations smarter.

• Style quiz identifies aesthetic preferences (60 seconds)
• Basic measurements for fit recommendations (60 seconds)
• First outfit generated immediately (30 seconds)
• Profile refines with every interaction

Results

Reflections

The hardest part isn't the AI—it's the product decisions. Which features actually matter to users? What's the right balance between personalization and simplicity? I'm learning that shipping fast and talking to users beats theorizing in isolation.