TrendSync Brand Factory

An end-to-end AI fashion factory: live trend intelligence, brand-compliant imagery, AI models wearing your designs, manufacturer-ready PDF tech packs, and Sora ad videos - all in one dashboard.

Comment

Inspiration

The fashion industry loses billions annually to trend misalignment - brands design collections months in advance, only to find they missed what consumers actually want. Independent designers face the same cycle: manually scrolling Instagram, guessing at trends, hoping their collections land. What if a single platform could analyze real-time global fashion trends, generate brand-compliant product imagery, dress AI fashion models in those products, produce manufacturing-ready tech packs as professional PDFs, and create ad-ready video - all from one dashboard? The "aha" moment was the runway video: upload your brand, and AI shows you your future collection on a runway.

What it does

TrendSync Brand Factory is an end-to-end AI fashion design platform where every capability is wired to the OpenAI model that does it best:

  • Trend Intelligence Engine - OpenAI Responses API with the hosted web_search tool analyzes real-time fashion trends across 6 global markets (LA, NYC, London, Tokyo, Paris, Seoul) - colors, silhouettes, materials, themes, celebrity influence - from live web data, cached in Redis with 24h TTL.
  • AI Collection Generator - GPT-5.6 SOL takes trend insights + brand guidelines through two-phase generation (plan the collection, then expand each product into 200-300 word image prompts) with validation and automated repair.
  • GPT Image generation and surgical editing - GPT-5.6 SOL writes art direction; gpt-image-2 renders and edits products. A surgical-vs-global classifier decides whether "change off-white to black" recolors one region or shifts the whole palette - covered by a regression test suite.
  • Company Models catalog + composites - 12 AI-generated fashion models (gpt-image-1.5) across diverse demographics; gpt-image-2 multi-image editing dresses a chosen model in your garment while preserving face, pose, and skin tone.
  • Lux, an Agents SDK design companion - built on the OpenAI Agents SDK (Python) with 7 typed tools: image analysis, surgical editing, brand-compliance adjustment, trend fetching, validation, variation, save-design.
  • Voice Design Companion - a Node.js microservice bridging the browser to OpenAI Realtime over WebSocket with raw binary PCM frames (no JSON+base64 overhead). The voice agent shares all design tools and adds ad-video generation, page navigation, and one-shot collection generation.
  • Tech packs + Foxit PDF pipeline - GPT-5.6 TERRA produces manufacturing specs (fabrics, measurements, construction, QC, packaging), persisted to Supabase as the single source of truth; python-docx builds styled DOCX and Foxit PDF Services converts, compresses, and merges them into tech packs and multi-product lookbooks.
  • Ad Video Generator (Sora 2) - GPT-5.6 TERRA plans a 5-scene storyboard (hook, hero, detail, lifestyle, CTA), then sora-2-pro renders each image-guided scene into a cinematic ad. The same pipeline makes single-product social clips.
  • Brand Guardian - a rule-based compliance engine (pure math, no AI): Euclidean RGB distance against the brand palette, camera/lighting range checks, negative-prompt scanning, with a transparent scoring formula per product.
  • Extras - Miro board export for production review, login auditing with real-time email alerts, and a demo account for instant judging.

How I built it

React + TypeScript + Vite frontend; Python backend services plus a Node.js voice microservice; Supabase for auth/data/storage; Redis for trend caching; OpenAI Responses API, Agents SDK, Realtime, gpt-image-2/1.5, and Sora 2 for the AI layer; Foxit PDF Services for the document layer; deployed on Vercel.

Challenges I ran into

  • Surgical image edits: distinguishing "recolor only the sole" from "rebrand the whole palette" required a dedicated classifier that understands preservation language, AI-rephrased instructions, and short voice phrasings.
  • Long structured outputs: two-phase collection generation with truncated-JSON repair and up to 3 retries to guarantee structurally complete collections.
  • Voice latency: switching the Realtime bridge to raw binary PCM WebSocket frames eliminated JSON+base64 overhead; Whisper language pinning stopped transcription drift.
  • PDF trust: tech packs are saved to the database BEFORE PDF generation is allowed, so the PDF always matches what the designer approved - never a re-hallucinated variation.

Accomplishments that I'm proud of

  • A complete loop: "what is trending right now?" to manufacturer-ready PDF and a cinematic ad video, in one dashboard.
  • Model composites that preserve identity - face, pose, skin tone - while swapping garments.
  • A voice companion that can drive the whole platform, including generating an ad video by asking for it.
  • Professional deliverables: styled tech packs and merged lookbooks that look agency-made.

What I learned

  • Route each job to the model that does it best; orchestration beats one-model-does-everything.
  • Deterministic guards (Brand Guardian, single-source-of-truth tech packs) are what make AI output trustworthy enough for manufacturing.
  • Real-time web search grounding beats stale trend datasets.

What's next for TrendSync Brand Factory

Runway video for full collections, marketplace integrations, collaborative team workspaces, and manufacturer handoff APIs.

Built With

Share this project:

Updates