Stilo
Mobile app that builds outfits from your own clothes using AI. Includes a colorimetry module that finds your most flattering palettes, a RAG pipeline that turns plain-language prompts into coordinated outfits, and an interactive canvas for drag-and-drop editing.
Co-Author & Lead Developer
The Problem
Closet overabundance increases daily cognitive load and leads to chronic underutilization of personal inventory. People gravitate toward the same few outfits while most garments sit idle. Stilo was built to test whether an AI system could reduce the mental effort of choosing what to wear and improve wardrobe rotation.
How It Works
- 1When you create your account, you fill out a style profile questionnaire and take a photo of your face. The app analyzes your facial colorimetry to find the color palettes that suit you best.
- 2You upload photos of your clothes or browse items from partner stores. The app removes backgrounds and extracts dominant colors and tags from each piece.
- 3You can assemble outfits manually on the interactive canvas, or describe the occasion in plain language and let the RAG pipeline generate coordinated outfits that respect your color palette and style preferences.
- 4You can iterate, 'make it more formal' or 'swap the jacket', and the system regenerates without starting over.
What I Built
Designed the three-phase RAG pipeline: a planning step where Gemini breaks prompts into clothing categories, a vector similarity search via pgvector to retrieve matching items from the wardrobe, and a composition step that assembles coherent outfits factoring in the user's color palette and style preferences.
Built the React Native mobile app with an interactive canvas for manually arranging outfits, drag, scale, rotate, plus the wardrobe and recommendation flows.
Provisioned all cloud infrastructure on Azure using Terraform.
Results
A controlled experiment with 10 participants compared manual vs. AI-assisted outfit creation using a standardized 25-item capsule wardrobe. The AI condition reduced UI interactions by 83.4% (p = 0.001), lowered cognitive load by 22.3% (NASA-TLX), and scored 86.75/100 on the System Usability Scale.