About Visual Vibe Check ✨

Inspiration I wanted to build something genuinely fun while diving deep into machine learning and dataset creation. Most personality quizzes feel formulaic and boring - I was inspired to create something that felt more like discovering a hidden part of yourself through the images that naturally draw your eye. The challenge of building real ML from scratch in a hackathon timeframe was exciting.

What It Does Visual Vibe Check analyzes your photo preferences in real-time to reveal your aesthetic personality. Users choose their favorite images from 8 rounds of randomly generated photos, and my custom ML algorithm extracts visual features (color temperature, contrast, symmetry, complexity) to determine personality combinations like "Electric Dragon" (vibrant, warm energy) or "Zen Vampire" (balanced, mysterious darkness). Each result includes personalized descriptions and shareable visual preferences.

How I Built It I used Bolt.new for rapid development with Claude for ideation and prompt engineering. The core challenge was building authentic ML workflows - I implemented computer vision analysis extracting 6+ visual dimensions from each image, then created a relative normalization algorithm that compares all images within each round to eliminate bias. After struggling with API limitations (Unsplash costs, Pexels CORS issues), I pivoted to Lorem Picsum with sophisticated feature analysis. The frontend uses React with responsive design optimized for both mobile and desktop.

Challenges I Ran Into Biased ML features were the biggest nightmare. Initially, certain personality traits (like "Nuclear" and "Dragon") dominated every result because my feature extraction algorithms were too sensitive. I spent hours debugging why users kept getting identical personalities, diving deep into console logs to discover that contrast and warmth features were maxing out at 0.99 for every image. The solution required implementing relative feature normalization - analyzing all 6 images together and scoring them against each other rather than absolute thresholds.

The second major challenge was avoiding hardcoded personality assignments while maintaining meaningful ML. I had to resist the temptation to manually assign images to personality types and instead build algorithms that genuinely discover visual patterns.

Accomplishments That I'm Proud Of I'm most proud of designing real ML workflows that actually work while keeping the experience genuinely fun and engaging. The relative normalization algorithm was a breakthrough - it guarantees personality variety without sacrificing authentic analysis. Creating 16+ unique personality combinations with witty, non-repetitive descriptions that people actually want to share feels like a major win. The responsive design scales beautifully from mobile to desktop without breaking.

What I Learned Building authentic ML is exponentially harder than using APIs, but infinitely more rewarding. I learned that bias in training data/feature extraction is insidious - you can have sophisticated algorithms that still produce meaningless results if your foundational assumptions are wrong. The debugging process taught me to always verify that your ML is actually discovering patterns rather than reinforcing built-in biases. Prompt engineering with Claude for iterative development was surprisingly effective for rapid prototyping.

What's Next for Visual Vibe Check I want to keep developing this! The immediate priorities are making results truly shareable (social media integration, beautiful result cards) and refining the UI to be more engaging with animations and micro-interactions. Longer-term, I'd love to expand the personality framework, add mood board generation from selected images, and potentially train on larger, more diverse image datasets. The foundation is solid - now it's about polish and virality.

Built With

• bolt.new
• claude