Product Genesis

Inspiration

What if launching a product was as simple as sending a text message?

Today, bringing a product from concept to marketing campaign requires an entire creative army—product designers sketching concepts, art directors refining aesthetics, photographers staging shoots, marketing strategists crafting campaigns. Each handoff introduces inconsistency. Each iteration burns weeks and thousands of dollars.

We looked at BRIA's FIBO API and saw something others missed: structured JSON generation isn't just a feature—it's the foundation for an entirely new workflow paradigm.

Traditional image generation is chaos for multi-step processes. You prompt, you hope, you manually iterate. But what if AI agents could pass perfect structured data to each other, like a relay team passing a baton?

That's when Design DNA was born—a JSON fingerprint that ensures your product looks identical whether it's on a billboard in Times Square or a thumbnail on Instagram.

We asked ourselves: Can 5 AI agents replace an entire creative department?

The answer is yes.

What it does

Input: "Water bottle for 10-year-old kids"

Output: 5 unique product designs, 1 quality-validated winner, and 3 market-ready marketing assets.

Here's the magic:

User types 8 words ↓ ┌─────────────────────────────────────┐ │ 🧠 Agent 1: Product Strategist │ → Analyzes requirements, generates 5 distinct concepts │ 🎨 Agent 2: Visual Designer │ → Converts concepts to FIBO images │ ✅ Agent 3: Quality Critic │ → Scores, ranks, rejects failures, triggers regeneration │ 📢 Agent 4: Marketing Strategist │ → Crafts billboard, social, poster campaigns │ 🖼️ Agent 5: Asset Producer │ → Generates all marketing assets with Design DNA └─────────────────────────────────────┘ ↓ Complete product lifecycle in minutes

The breakthrough: Every marketing asset shares the same Design DNA. The water bottle in your billboard is pixel-perfect identical to the one in your Instagram post. No manual alignment. No "close enough." Perfect consistency, automatically.

Zero prompt engineering required. You don't need to know about camera angles, lighting setups, or color temperatures. The agents handle all of that through structured JSON communication.

How we built it

Architecture

We built a 5-agent pipeline where each agent has a specialized role:

| Agent | Role | Output | |----------------------|--------------------------------------------------|------------------------------------------| | Ideation Specialist | Analyzes user intent, generates diverse concepts | 5 Product Concept JSONs | | Visual Designer | Translates concepts to FIBO parameters | 5 Photorealistic Images | | Quality Validator | Scores, ranks, enforces brand compliance | Winner Selection or Regeneration Request | | Marketing Strategist | Creates campaign concepts | Billboard, Social, Poster Specs | | Asset Generator | Produces final marketing materials | Market-Ready Assets |

Tech Stack

Frontend: Next.js 14 + React 18 + Tailwind CSS Backend: Python FastAPI + Async Orchestration LLM: Ollama (Llama 3.2) - runs locally Images: BRIA FIBO API - JSON-native generation Realtime: WebSocket for live agent updates Schemas: Pydantic for type-safe JSON validation

The Pipeline Flow

# Simplified orchestration logic async def run_pipeline(user_prompt: str): # Stage 1: Ideation concepts = await ideation_agent.process(user_prompt)

  # Stage 2: Design Generation
  designs = await design_agent.process(concepts)


  # Stage 3: Quality Validation (with regeneration loop)
  for attempt in range(MAX_RETRIES):
      validation = await validator_agent.process(designs)
      if validation.approved:
          break
      designs = await design_agent.regenerate(validation.feedback)


  # Stage 4: Marketing Strategy
  marketing = await marketing_agent.process(validation.winner)


  # Stage 5: Asset Generation (preserving Design DNA)
  assets = await asset_agent.process(marketing, design_dna=validation.winner)


  return assets

Design DNA Implementation

The core innovation is preserving product identity across assets:

{ "design_dna": { "product_type": "water_bottle", "name": "Adventure Buddy", "colors": {"primary": "#FF6B35", "secondary": "#1E90FF"}, "material": "BPA-free plastic with soft-touch grip", "style": "playful_modern" } }

This DNA propagates to every marketing asset, ensuring the product in your billboard matches the product in your social post.

Challenges we ran into

The Consistency Problem

Challenge: Making the same product appear identical across billboard, poster, and social media seemed impossible with traditional prompting.

Solution: Design DNA—a core JSON fingerprint inherited by every marketing asset. FIBO's deterministic output from structured parameters made this possible.

The Quality Control Paradox

Challenge: How do you programmatically judge if an AI-generated image is "good enough"?

Solution: We built a multi-criteria scoring system:

$$\text{Overall Score} = \frac{\sum_{i=1}^{6} w_i \cdot s_i}{\sum_{i=1}^{6} w_i}$$

Where criteria include requirement alignment, visual quality, target appeal, brand potential, uniqueness, and brand compliance. Designs scoring below 70/100 trigger automatic regeneration with specific JSON corrections.

The Prompt Engineering Trap

Challenge: Early versions still required careful, detailed prompts to get good results.

Solution: We rewrote agent system prompts iteratively until a literal 8-word input ("water bottle for 10-year-old kids") produces professional-grade outputs. The agents infer everything else.

Agent Communication Overhead

Challenge: Getting 5 agents to collaborate without losing context or producing inconsistent outputs.

Solution: Strict Pydantic schemas at every handoff point. Agents don't pass vague messages—they pass validated JSON structures.

Accomplishments that we're proud of

🏆 Zero to Campaign in Minutes — What traditionally takes weeks now happens in minutes. Type 8 words, get market-ready assets.

🏆 Design DNA Concept — We invented a new primitive for AI-generated content: a JSON fingerprint that maintains identity across unlimited derivatives.

🏆 Self-Healing Pipeline — Our validation agent doesn't just score—it fixes. Brand compliance violations trigger automatic corrections and regeneration.

🏆 True JSON-Native Workflow — Not a single vague prompt in the entire pipeline. Every agent communicates through structured, type-safe JSON.

🏆 Real-Time Transparency — Users watch each agent think in real-time via WebSocket. It's not a black box—it's a glass box.

🏆 No Prompt Engineering Required — We've hidden all complexity. Users describe what they want in plain English; agents handle the translation to professional FIBO parameters.

What we learned

Technical Insights

JSON > Prompts for Multi-Step Workflows — Vague prompts compound errors across pipeline stages. Structured JSON eliminates drift.
Validation is the Secret Sauce — Most AI tools generate and hope. Adding a critic agent that can reject and regenerate transforms output quality.
FIBO's Disentanglement is Underrated — The ability to change lighting without affecting the product, or adjust colors without breaking shape, is what makes Design DNA possible.

Process Learnings

Agent Specialization Matters — One mega-agent performing all tasks produces worse results than 5 specialists collaborating.
Fail Fast, Regenerate Faster — Building in retry loops with specific feedback actually speeds up the pipeline because it prevents bad outputs from propagating.

What's next for FPLE

Short-Term Roadmap

🎨 Interactive Refinement Panel — Let users tweak FIBO parameters (colors, lighting, angles) in real-time and see instant regeneration
📦 Template Library — Pre-built workflows for common product categories (beverages, electronics, fashion, cosmetics)
🔄 Batch Processing — Generate campaigns for entire product catalogs in one run

Long-Term Vision

🌐 Brand Asset Platform — Upload your brand guidelines once, and every asset generated automatically complies
📊 A/B Testing Integration — Generate multiple campaign variants and feed performance data back to improve future generations
🤖 Agent Fine-Tuning — Train specialized agents on industry-specific aesthetics (luxury, kids, eco-friendly, etc.)

The Dream

We envision a world where any entrepreneur can go from product idea to market-ready campaign in an afternoon. No design team. No agency fees. No weeks of iteration.

Just your idea, FIBO, and 5 AI agents working together.

One prompt. Five agents. Infinite possibilities.