Membangun AI Fashion Stylist dari Nol — Behind the Scenes OutfittingMe

Di artikel ini, gue mau share cerita di balik pembangunan OutfittingMe — platform AI fashion stylist pertama di Indonesia. Bukan cuma soal teknologi, tapi juga soal product decisions, engineering trade-offs, dan lessons learned.

The Problem Statement

Indonesia punya pasar fashion e-commerce senilai $6.2 miliar. Tapi ada gap besar: 30% pembelian online dikembalikan karena ga cocok. Kenapa? Karena ga ada tools yang bener-bener memahami bentuk tubuh dan preferensi personal.

Existing solutions? Stylist manusia mahal ($500-5000/sesi). Recommendation engine marketplace? Cuma berdasarkan "orang yang beli X juga beli Y" — ga memahami body shape atau color matching.

Architecture Overview

OutfittingMe dibangun dengan stack modern yang optimized untuk AI processing:

Backend: Fastify 5 (Node.js) — fast, lightweight, excellent TypeScript support
Frontend: Next.js 15 — SSR untuk SEO, React untuk interactivity
Database: PostgreSQL 16 dengan pgvector extension — untuk vector similarity search
AI/ML: Ollama embeddings (768D), GPT-4o-mini untuk VTO
Cache: Redis 7 — session management dan recommendation caching
ORM: Prisma — type-safe database access

The 9-Layer Recommendation Engine

Core innovation dari OutfittingMe adalah 9-layer recommendation engine. Kenapa 9 layer? Karena fashion itu complex — ga cukup cuma "kamu suka warna biru, ini produk biru".

Setiap layer memproses data secara berurutan:

Body Shape Analysis — Input: measurements → Output: shape classification (pear/apple/hourglass/rectangle/inverted triangle)
Undertone Detection — Input: skin analysis quiz → Output: warm/cool/neutral
12-Season Color Harmony — Input: undertone + hair/eye color → Output: seasonal palette
Silhouette Scoring — Input: body shape + garment shape → Output: compatibility score 0-100
Textile Physics — Input: fabric type + body type → Output: draping prediction
Modest-Fit Scoring — Input: modest requirements + garment → Output: modesty score
Style Personality — Input: quiz answers → Output: style archetype
Occasion Matching — Input: context + garment → Output: appropriateness score
Budget Optimization — Input: budget + options → Output: ranked list

Vector Search for Fashion

Salah satu technical challenge terbesar adalah product matching. Dengan 10,000+ produk, gimana cara cari yang paling cocok dalam milliseconds?

Jawabannya: pgvector. Setiap produk di-embed ke 768-dimensional vector menggunakan Ollama. Ketika user request recommendation, profil mereka juga di-embed, lalu cosine similarity search di PostgreSQL. Result: sub-100ms response time.

Virtual Try-On Implementation

VTO Studio menggunakan GPT Image generation untuk bikin photorealistic outfit visualization. Challenge-nya:

Latency: Image generation butuh 5-15 detik. Solution: streaming response + progress indicator
Quality vs Speed: Dua mode — Speed (1 credit, 2 items) dan Quality (2 credits, 6 items)
Body Accuracy: Manekin virtual harus representatif. Solution: parameterized body model dari measurements

Lessons Learned

Start with data: Fashion AI butuh data yang banyak. Kami mulai dengan affiliate integration untuk populate product catalog
User feedback loop: Setiap recommendation yang user like/skip = training data untuk improve model
Cultural sensitivity: Modest fashion punya requirement yang unik. Generic AI models ga handle ini — perlu custom scoring
Performance matters: Recommendation harus real-time. Pre-compute yang bisa, cache aggressively

Tech Stack Decision Log

Fastify over Express: 2-3x faster, built-in TypeScript, schema validation
PostgreSQL over MongoDB: ACID compliance untuk transactions, pgvector untuk vector search
Prisma over TypeORM: Better TypeScript integration, migration system, studio untuk debugging
Ollama over OpenAI embeddings: Self-hosted, no API cost, 768D quality sufficient
Next.js over plain React: SEO critical untuk fashion platform, SSR mandatory

What's Next

OutfittingMe masih dalam early stage. Roadmap:

Brand partner dashboard (live analytics untuk brand)
AR try-on (real-time camera overlay)
Social features (share outfits, community voting)
API access untuk third-party integration

Interested in the tech? Check out outfittingme.com atau reach out ke [email protected].