TrendSync Brand Factory

AI-powered fashion design platform that analyzes real-time trends via Gemini 3 + Google Search, generates brand-compliant collections with AI imagery, and offers a voice design companion.

Comment

TrendSync Brand Factory

AI-powered fashion design platform that analyzes real-time trends via Gemini + Google Search, generates brand-compliant collections with AI imagery, produces manufacturing-ready tech packs as PDFs via Foxit, creates ad videos with Veo 3.1, and offers both typing and voice AI design companions — all powered by Google's Gemini 3 ecosystem and deployed serverless on Google Cloud Run.

Live Demo: trendsync-brand-factory.vercel.app

Inspiration

The fashion industry loses billions annually to trend misalignment — brands design collections months in advance, only to find they've missed what consumers actually want. I watched independent designers struggle with the same cycle: manually scrolling Instagram, guessing at trends, and hoping their collections would land. I thought — what if AI could close that gap entirely? What if a single platform could analyze real-time global fashion trends, generate brand-compliant product imagery, produce manufacturing-ready tech packs as professionally formatted PDFs, and even create ad-ready video content — all powered by Gemini and enhanced by Foxit's document automation?

What it does

TrendSync Brand Factory is an end-to-end AI fashion design platform built entirely on Google's Gemini 3 ecosystem with Foxit document automation for professional output:

  • Trend Intelligence Engine: Uses gemini-2.5-flash with Google Search grounding to analyze real-time fashion trends across 6 global markets (LA, NYC, London, Tokyo, Paris, Seoul), broken down by colors, silhouettes, materials, themes, and celebrity influence — all sourced from live web data, not stale datasets. Includes automatic retry logic for rate limits and truncated JSON repair when Gemini responses are cut off mid-stream. Results are cached in Redis with 24-hour TTL to minimize API costs.

  • AI Collection Generator: Takes trend insights + brand guidelines and uses gemini-3-pro-preview with thinking_level=HIGH for deep multi-step collection planning. A two-phase approach — Phase A plans the collection structure, Phase B expands each product with 200-300 word image prompts — followed by validation and automated repair (up to 3 retries) ensures every collection is structurally complete.

  • AI Image Generation: A two-step pipeline — gemini-2.5-flash builds a detailed art-direction prompt incorporating brand style, lighting, camera settings, and trend data, then gemini-3-pro-image-preview (Gemini's native image generation model) generates the actual product image. Images are stored in Google Cloud Storage with signed URLs, falling back to base64 data URLs when needed.

  • Brand Guardian: A rule-based compliance engine that scores every generated product against the brand's color palette, negative prompts, camera settings, and lighting configuration — automatically flagging violations. Compliance scores are stored per product and displayed in the gallery view.

  • Design Companion Chat ("Lux"): A Google ADK agent powered by gemini-2.5-flash with 7 specialized tools — image analysis, image editing, brand compliance adjustment, trend data fetching, compliance validation, image variation generation, and design saving. Built on Vertex AI with a critical architectural innovation: large data (base64 images) is stored externally in _IMAGE_STORE to prevent ADK's history serialization from exceeding token limits. Includes intelligent image compression (images > 500KB are resized to max 1024px JPEG before editing) and automatic 429 retry logic with exponential backoff. Fresh sessions per request prevent history accumulation.

  • Tech Pack Generator + PDF Pipeline: Uses gemini-3-pro-preview to produce manufacturing-ready technical specifications — fabric details, measurements, construction notes, quality control standards, packaging — from a single product description. Tech packs are persisted to Supabase as the single source of truth: the techpack_json column with a techpack_generated flag ensures that PDFs always reflect exactly what the designer approved in the UI, never a re-hallucinated variation. PDFs are generated via a python-docx → Foxit Cloud pipeline (see dedicated section below).

  • Voice Design Companion ("Lux"): A Google ADK agent using gemini-live-2.5-flash-native-audio that lets designers talk through design decisions hands-free via bidirectional WebSocket streaming — real-time PCM audio at 16kHz as binary WebSocket frames (no JSON+base64 overhead) with live status feedback ("Editing image...", "Fetching trends..."). Both typing and voice agents share the same 7 tools (shared/design_tools.py) and image pipeline, ensuring identical capabilities regardless of input modality. The voice agent additionally has 3 exclusive tools: video generation, page navigation, and collection generation.

  • Full Pipeline Orchestration: A single endpoint (POST /adk/pipeline) chains the entire workflow: trends → collection plan → product image generation → ad video — executing all four steps as a coordinated pipeline with status polling.

  • Login Monitoring: Every sign-in is double-logged — a login_audit record in Supabase (user ID, browser, timestamp) plus a real-time email notification via Resend API to the admin, providing instant awareness of platform activity.

How Foxit Powers Professional Document Output

Foxit's document automation APIs are central to TrendSync's professional output pipeline. We use Foxit PDF Services for document conversion, compression, and merging — creating a complete "generate, process, deliver" workflow that turns AI-generated fashion data into manufacturer-ready documents.

The Problem Foxit Solves

AI can generate brilliant fashion collections, but the fashion industry runs on PDFs. Manufacturers need tech packs. Buyers need lookbooks. Emails need attachments. Without professional document output, an AI platform is just a demo. Foxit bridges the gap between AI intelligence and industry-standard deliverables.

Brand Guardian — How Validation Works

The Brand Guardian is a rule-based compliance engine (not hardcoded scores). It validates every product's design specification against the brand's style configuration in real-time.

Validation Checks

Check What It Does Severity
Color Palette Extracts hex colors from the product's color_scheme and measures Euclidean RGB distance against the brand palette. If distance > 30 (perceptually different), it flags a violation. suggestion
Camera Settings Checks focal length (converted to FOV) and camera angle against brand-defined min/max ranges. warning
Lighting Compares lighting temperature (warm vs cool) against the brand's configured color temperature (e.g., 5000K). suggestion
Negative Prompts Scans product description and object descriptions for forbidden terms defined in brand style (e.g., "blurry", "low quality"). critical

Scoring Formula

compliance_score = 100 - (critical x 25) - (warning x 10) - (suggestion x 3)
  • 100% = No violations found — product fully matches brand guidelines
  • 75-99% = Minor suggestions (e.g., trend colors differ from brand palette)
  • 50-74% = Warnings present (e.g., camera angle out of range)
  • <50% = Critical violations (e.g., forbidden terms in description)

Where Brand Rules Are Stored

Brand style rules are stored in the Supabase brand_styles table as a JSONB column (style_json), configured via the Brand Style Editor page:

{
  "colorPalette": [{ "name": "Brand Navy", "hex": "#1a237e", "designation": "primary" }],
  "cameraSettings": { "fovMin": 20, "fovMax": 80, "angleMin": 0, "angleMax": 90 },
  "lightingConfig": { "colorTemperature": 5000 },
  "negativePrompts": ["blurry", "low quality", "distorted"],
  "materialLibrary": [...],
  "logoRules": {...}
}

Implementation: trendsync-backend/shared/brand_guardian.py (validate_prompt() function)

How I built it

The architecture is a three-tier system designed to keep Gemini at the center of every intelligent decision, deployed serverless on Google Cloud Run with Foxit handling the professional document output layer:

Frontend — React 18 + TypeScript + Vite with a custom neumorphic pastel design system, deployed on Vercel. Every AI interaction goes through a centralized API client (api-client.ts) — zero direct Gemini calls from the browser, keeping API keys secure server-side.

Backend (3 FastAPI microservices on Google Cloud Run):

  • Main Backend: The brain. All Gemini calls route through here using google-genai SDK with Vertex AI (vertexai=True). Hosts the Google ADK design companion agent ("Lux"), endpoints for trends (Google Search grounding), collection generation, image gen/edit, tech packs, PDF generation via Foxit, lookbook export, ad video orchestration, and a WebSocket proxy to the voice service. Includes Redis caching (24h TTL) to reduce API costs.
  • Video Generation Service: Dedicated Veo 3.1 pipeline for the "Future You" ad video feature. Takes collection data + product images, generates cinematic fashion videos in us-central1.
  • Voice Companion: WebSocket server running a Google ADK agent with gemini-live-2.5-flash-native-audio for real-time bidirectional voice interaction during design sessions. Features tool status feedback so designers see what's happening during long operations.

All three services share a single Docker image with an entrypoint.py that dynamically loads the correct service via the SERVICE env var — deployed to Cloud Run with min-instances=1 on the main backend (eliminating cold starts) and auto-scale-to-zero on video and voice services.

Document Layer — Foxit PDF Services API for DOCX-to-PDF conversion, compression, and multi-document merging. python-docx generates styled DOCX files locally; Foxit's cloud handles the rest.

Database & Auth — Supabase (PostgreSQL + Auth) with 11 tables, Row-Level Security, a Redis caching layer for trend insights and structured prompts, and a login_audit table for security monitoring.

Cloud Storage — Google Cloud Storage for product images and generated videos, with signed URLs for secure access.

Email — Resend API for login notification emails (admin alerts on every sign-in) and tech pack email delivery with PDF attachments.

trendsync Image trendsync Image trendsync Image trendsync Image

Key Gemini integration points:

Task Model Notes
Trend analysis gemini-2.5-flash Google Search grounding, Redis cache, truncated JSON repair
Collection planning gemini-3-pro-preview thinking_level=HIGH, 2-phase + repair
Image prompt building gemini-2.5-flash Art direction with brand style
Product image generation gemini-3-pro-image-preview Two-step pipeline, compression
Tech pack generation gemini-3-pro-preview Structured output, Supabase persistence
Design Companion gemini-2.5-flash ADK agent, 7 tools, external image store
Voice Companion gemini-live-2.5-flash-native-audio ADK agent, BIDI streaming, 10 tools
Ad video storyboard gemini-3-pro-preview 5-scene storyboard, HIGH thinking
Video generation veo-3.1-generate-preview Cinematic clips, style reference images

The critical architectural decision was using Google Search as a grounding tool for trend analysis — this gives TrendSync access to live fashion data rather than training cutoff knowledge, making every trend report current to the day.

Architecture

React (Vercel) → Cloud Run: Main Backend (:8080) → Gemini 3 Pro / Flash / Image / Veo / Foxit / GCS
                                                  → Cloud Run: Video Service (:8080) for Veo 3.1
                                                  → Cloud Run: Voice Service (:8080) for Gemini Live audio
                          ↕
                     Supabase (PostgreSQL + Auth) + Redis Cache

Gemini Models Used

Model Location Purpose
gemini-3-pro-preview global Collection planning, tech packs, storyboards
gemini-3-pro-image-preview global Product image generation and editing
gemini-2.5-flash global Trend engine, art-direction prompts, ADK design companion
gemini-live-2.5-flash-native-audio — Voice design companion (bidirectional streaming)
veo-3.1-generate-preview us-central1 Ad video generation (5-scene cinematic)

trendsync Image

Challenges I ran into

Google Search grounding vs. structured JSON output: I discovered that google_search grounding is incompatible with response_mime_type="application/json" in the Gemini API. I had to implement a two-pass approach — first call with grounding enabled (free-form text enriched with live web data), then parse the grounded response into structured trend data with robust JSON extraction.

ADK token overflow from base64 images: Google ADK serializes every function_response into the conversation history for subsequent model calls. When a tool returned a base64-encoded image (~1.5MB of text), the accumulated context exploded past Gemini's token limit by the second edit. The solution: store large data in an external _IMAGE_STORE dictionary (never in the tool response), and extract images after run_async() completes. This pattern is critical for any ADK agent handling binary data.

Image edits reverting in the UI: After the agent edited an image, it would flash the new version and immediately revert. Root cause: two independent setInterval polling loops were fetching from the database every 2 seconds and overwriting the in-memory edited image (which hadn't been saved to DB yet). Removed both polling loops — in-memory state is now the source of truth for unsaved edits.

Gemini truncating JSON responses: Long collection plans would get cut off mid-JSON, breaking the parser. Built a _repair_truncated_json() function that closes open strings, removes trailing partial entries, and properly terminates arrays/objects — recovering ~90% of otherwise-lost responses.

429 rate limits not retrying: Gemini's RESOURCE_EXHAUSTED errors were shown as user-facing messages instead of being retried. Added two layers of retry with exponential backoff: 3 retries at the design_tools.py layer (8s/16s/24s) and 3 retries at the image_generator.py layer (5s/10s/15s) — effectively 9 attempts before giving up.

Slow image edits (15-25s to 8-12s): Images were 1.2MB PNGs being sent to Gemini for every edit, with each output slightly larger than the input. Added _compress_for_edit(): images > 500KB are resized to max 1024px and converted to JPEG (quality 85), giving a 6-8x size reduction. The 5th edit is now as fast as the 1st.

Gemini 3 API differences: Gemini 3 uses thinking_level (HIGH/MEDIUM/LOW enum), NOT thinking_budget (that's Gemini 2.5). Also requires location=global on Vertex AI, while Veo requires us-central1 — mixing these up causes silent failures.

Accomplishments that I'm proud of

Full pipeline from trend to video in one click: A single POST /adk/pipeline endpoint chains trend analysis → collection planning (Gemini 3 Pro with deep thinking) → product image generation (gemini-3-pro-image-preview) → 5-scene ad video (Veo 3.1). A designer can go from "what's trending?" to "here's my manufacturing-ready collection with professional tech pack PDFs and a promotional video" in a single session.

Two AI design companions sharing one brain: Both the typing agent and voice agent ("Lux") are Google ADK agents sharing the same 7 tools via shared/design_tools.py. A designer can type "make the jacket terracotta" or say it out loud — same tool executes, same image pipeline runs, same result. The voice agent is actually faster (~8-12s vs ~12-18s for image edits) because Gemini Live's persistent bidirectional session eliminates the routing roundtrip.

Real-time trend intelligence that actually works: The combination of gemini-2.5-flash + Google Search grounding produces trend reports that match what I see on Vogue and WGSN — colors, silhouettes, materials, themes — all grounded in current web data. This isn't hallucinated fashion advice; it's AI-synthesized market intelligence.

Professional document output via Foxit: The Foxit integration transforms AI-generated data into documents that look like they came from a professional design agency. Navy branded headers, structured measurement tables, compressed PDFs ready for email — this is the difference between a hackathon demo and a production tool. The lookbook merge feature (combining multiple tech packs into a single PDF) is something manufacturers actually need.

Single source of truth architecture: The tech pack persistence pattern — save to DB once, generate PDF from saved data only, clear on design changes — eliminates the #1 risk of AI-generated documents: inconsistency. What you see in the UI is exactly what appears in the PDF.

What I learned

Gemini 3's native image generation changes everything. Using gemini-3-pro-image-preview instead of external image APIs means the model that understands your design intent is the same model generating the image — no prompt translation layer, no API mismatch. The two-step pipeline (Flash writes art-direction prompt → Image model renders) gives precise brand-compliant results.

Google ADK is the right abstraction for multi-tool agents. Building the design companion with ADK's Agent + Runner + ToolContext pattern meant I could add tools incrementally, share them between typing and voice agents, and let Gemini handle tool selection — no manual routing logic. Fresh sessions per request prevent token accumulation.

Gemini's Google Search grounding is a game-changer for real-time applications. Most AI apps are limited by training cutoffs — TrendSync breaks that barrier entirely. Fashion trends shift weekly; grounded search means the platform is always current.

gemini-2.5-flash is remarkably capable for production workloads. I expected to need Pro for most tasks, but Flash handled trend analysis, conversational design chat, tech pack generation, and brand compliance reasoning with excellent quality at a fraction of the cost and latency. I reserved gemini-3-pro-preview only for the most complex multi-step reasoning tasks (collection planning, storyboards).

The Gemini ecosystem is genuinely composable. Gemini 3 Pro for deep reasoning, Gemini 3 Pro Image for generation, Flash for speed, native audio for voice, Veo 3.1 for video, Google Search for grounding, Google ADK for agent orchestration — these aren't separate products bolted together; they share the same SDK patterns (google-genai with vertexai=True) and work naturally as a unified AI backend.

Architecture matters more than model size. The biggest improvements came not from switching models, but from designing the right prompts, caching strategy (Redis with 24h TTL), image compression pipeline, retry logic, and data flow. A well-structured gemini-2.5-flash call with proper context outperformed naive Pro calls every time. Similarly, the single source of truth pattern for tech packs solved a consistency problem that no amount of prompt engineering could fix.

What's next for TrendSync Brand Factory

  • Multi-brand portfolio management: Supporting agencies that manage multiple fashion brands, each with distinct guidelines, from one dashboard.
  • Supplier matching: Using Gemini to match tech pack specifications with a database of global manufacturers, completing the design-to-production pipeline.
  • Mobile companion app: A voice-first mobile interface using gemini-live-2.5-flash-native-audio so designers can iterate on collections while away from their desk — "Hey TrendSync, swap the jacket color to the trending terracotta I saw in the Seoul report."

Built With

  • geminiapis3
  • google-agents-sdk
  • google-gravity
Share this project:

Updates