Highlights
Project snapshot
- Retrieval-grounded ideation: trends embedded with SentenceTransformers (384-dim) and matched via pgvector RPCs.
- Synthetic-persona validation with Microsoft TinyTroupe (1–500 AI participants) before any human panel.
- Queue-backed generative imaging across Ideogram, Gemini, Replicate, and fal.ai via BullMQ/Redis workers.
Overview
Enterprise AI for product innovation teams
Creator AI is a multi-tenant SaaS that guides enterprise innovation teams — primarily FMCG / consumer-goods brands — through the full product-ideation lifecycle: scraping live market data, grounding ideas in retrieved trends, generating concepts and photoreal packshots, and stress-testing them against AI-simulated audiences.
It is a Next.js application backed by a fleet of Python micro-services, a Redis/BullMQ job system, and a Supabase (Postgres + pgvector) data layer with company-scoped access control.
Problem
Research and ideation were fragmented and low-context
- Market research, trend analysis, concept creation, and visual production lived in separate tools and teams.
- Concept validation traditionally required slow, expensive human consumer panels.
- Generating brand-consistent product imagery at scale was a manual creative bottleneck.
- Enterprise rollout demanded multi-tenant isolation, role-based controls, and auditable AI usage.
Solution
A structured workflow with retrieval, generation, and governance
- A guided Briefing → Ideation → Generation flow per company, brand, and project.
- RAG over a pgvector trend store, synthetic-persona testing with TinyTroupe, and multi-provider image generation in one pipeline.
- Company-scoped multi-tenancy, admin role gating, and step-level LLM instrumentation for governance.
Core Features
Capabilities for enterprise product teams
Guided Innovation Workflow
- Per-company brands, projects, data silos, and looks
- Drag-and-drop concept management
- Per-company run tracking
Research & Retrieval (RAG)
- Live scraping via Crawl4AI, Firecrawl, and Scrapy/Playwright spiders
- SentenceTransformers 384-dim embeddings matched through pgvector RPCs
- Search-volume enrichment to augment trend scoring
Synthetic-Persona Validation
- TinyTroupe agent panels (1–500 configurable personas)
- Automated concept reactions before human research
- Pydantic-typed product ideas with thread-safe concurrency
Creative Asset Generation
- Multi-provider imaging: Ideogram, Gemini "Nano Banana", Replicate, fal.ai
- OpenAI text models (incl. o3-pro) for ideation, copy, and grading
- PDF export of concepts and an OpenAI Realtime voice console
Architecture
Next.js APIs backed by queues, storage, and Python services
Next.js (App Router) UI + API
→ BullMQ queues (idea / packshot / brandAssets) over Redis (TLS)
→ standalone Node/TSX worker (LLM + image pipeline)
→ FastAPI services (Crawl4AI / research+TinyTroupe / search-volume)
→ Supabase Postgres (pgvector) + Supabase Storage (signed URLs)
- Three BullMQ queues with exponential backoff, removeOnComplete/Fail, and stalled-job recovery decouple slow generative work from request/response.
- A polyglot Scrapy + scrapy-playwright + Zyte spider fleet feeds fresh market data across many retail sites.
Technical Highlights
Vector RAG, durable queues, and multi-provider AI
- Vector RAG over pgvector with a strict 384-dim embedding contract and Postgres similarity RPCs.
- Multi-provider AI abstraction composing OpenAI (o3-pro + Realtime), Google GenAI, Ideogram, Replicate, fal.ai, and HuggingFace behind one job processor.
- Supabase-auth SSR with middleware token refresh, is_admin / is_active gating, and company-scoped multi-tenancy.
- An LlmFlowTracker instruments every LLM step (duration, provider, model) for cost and performance observability.
Business Value
More concepts, lower cost, enterprise-governed
- Compresses a multi-team, multi-week innovation cycle into a single governed platform.
- Lets teams explore far more concepts at a fraction of traditional research and creative-production cost.
- Keeps AI usage observable, role-controlled, and tenant-isolated for enterprise deployment.