P'STYLST

P'STYLST is an intelligent digital platform that democratizes personal styling. It was built to solve a fundamental disconnect in the fashion industry: the gap between knowing what clothes look good and knowing what clothes look good on you.

By combining computer vision, body analysis, and advanced LLMs, P'STYLST provides bespoke wardrobe intelligence that was previously only accessible via high-end personal stylists. This matters because confidence is inherently linked to personal presentation, and technology should empower individuals to look and feel their best.

Fashion e-commerce suffers from a massive returns problem (often exceeding 30%) because users buy clothes that don't suit their body type or skin tone.

• Current Situation: Users rely on generic size charts and model photos.
• Pain Points: Decision paralysis, buyer's remorse, and wasted money on clothes that don't fit well.
• Market Gap: No affordable, personalized, data-driven styling solution exists for the average consumer.
• Business Challenge: High return rates destroy e-commerce profit margins.

Mission: To eliminate wardrobe anxiety through personalized data intelligence.

Vision: A world where every individual has a trusted digital stylist in their pocket, ensuring they only buy what truly complements them.

North Star Metric: Confidence Score (measured by user survey post-recommendation) and Outfit Generation retention rate.

Business Model: Freemium access to basic body analysis, with a premium subscription for unlimited wardrobe generation and occasion-specific styling.

MVP Strategy: Focus heavily on the accuracy of the computer vision body analysis first. If the foundation is wrong, the LLM recommendations will be useless. Prioritize a mobile-first web experience to leverage device cameras easily.

• Functional: Real-time body scanning, wardrobe digitization, outfit generation based on weather/occasion.
• Non-functional: Sub-2-second latency on initial AI generation. Mobile-first responsiveness.
• AI Requirements: High accuracy in edge-based facial/body landmark detection. Guardrails against inappropriate image uploads.

AI is the brain of P'STYLST. I chose to use a hybrid approach to balance performance, cost, and privacy.

• Computer Vision (Edge): Used TensorFlow.js for immediate, private body and face analysis on the user's device. No images are sent to the server for this step.
• Recommendation System (Server): Integrated OpenAI's LLMs via structured outputs to generate semantic clothing recommendations based on the edge-computed physical data.
• Guardrails: Implemented strict moderation endpoints and fallback logic to handle edge cases where the CV model fails to detect landmarks accurately.

Why Next.js? SSR was critical for perceived performance, and the app router provided seamless integration between frontend UI and backend AI routes.

Trade-offs: Running TensorFlow.js on the client increases the initial JS bundle size. I mitigated this by dynamically importing the heavy CV modules only when the user enters the scanning flow, keeping the landing page incredibly fast.

• Short-term: Implement advanced caching for generated outfits to reduce LLM API costs.
• Mid-term: Direct e-commerce integrations to allow users to buy generated outfits immediately.
• Long-term: AR try-on capabilities directly in the browser.

Building P'STYLST fundamentally shifted my perspective on product engineering. It taught me that an AI model is only as valuable as the user experience surrounding it. The biggest learning was managing state between heavy client-side ML models and server-side LLMs gracefully.