Multi-market AI content production pipeline — from product brief to platform-ready copy in 12 languages.
Portfolio demo. All product briefs, brand names, customer references, and URLs in this repo are fictional. The pipeline architecture, code, and fixtures demonstrate the system — they are not tied to any real company or product.
ContentFactory accepts a product brief and brand voice profile, then fans out across up to 12 markets in parallel to produce localised, platform-ready marketing copy.
Each output goes through a full quality pipeline before it ever reaches a human reviewer:
Product Brief
│
▼
Claude (EN-US generation)
│
├─── DeepL ──────────────────┐
│ ▼
└───────────────── Claude Transcreation (cultural adaptation)
│
Back-translation QA
(keyword overlap < 40% = flag)
│
LLM-as-Judge Eval
(tone · brand voice · cultural · hallucination)
│
┌─────score ≥ 75──────┐
▼ ▼
Human Review Queue Auto-rejected
│
Approve / Reject
│
Shopify JSON · WordPress/WPML JSON · Export bundle
| Feature | Detail |
|---|---|
| 12 locale support | en-US, en-GB, fr-FR, fr-CA, es-ES, es-MX, de-DE, it-IT, pt-BR, nl-NL, pl-PL, ja-JP |
| 4 content types | Product descriptions · Ad copy · Meta tags · Landing page copy |
| 3 translation modes | deepl · claude · both (MT draft → Claude refine) |
| Inngest fan-out | N locales × M content types jobs run fully in parallel |
| Back-translation QA | Keyword overlap ratio detects semantic drift without an NLP service |
| LLM-as-Judge eval | 0–100 score across tone, brand voice, cultural accuracy, hallucination |
| HITL review queue | SLA-coloured review cards (green <24h · amber 24-48h · red >48h) |
| Fixture mode | All adapters have fixture implementations — zero API keys to run locally |
| Shopify output | product.metafields + shopify_translation adapter-shaped JSON |
| WordPress output | Post body · Yoast SEO meta · WPML hreflang map |
| Eval analytics | Recharts radar + bar charts aggregated by locale and content type |
| Export bundle | Single-click project export as structured JSON |
flowchart TD
A([📋 Product Brief\nBrand Voice Profile]) --> B[/POST /api/pipeline/generate/]
B --> C[(Supabase\nprojects · markets)]
C --> D{{"🎯 Inngest\ncontentfactory/project.submitted"}}
D --> E["⚡ generateContentFunction\nCreate N×M jobs · fan-out"]
E --> F1{{"job.started en-US×desc"}}
E --> F2{{"job.started fr-FR×copy"}}
E --> FN{{"… N×M total"}}
F1 & F2 & FN --> G["translateMarketFunction\n(parallel workers)"]
G --> H["Claude · DeepL\nGenerate · Translate · Transcreate"]
H --> M["LLM-as-Judge Eval\n0–100 score"]
M -->|"≥ 75"| N["✅ Human Review Queue"]
M -->|"< 75"| O["❌ Auto-rejected"]
N --> P["👤 HITL Approve/Reject"]
P --> Q["Shopify JSON · WordPress JSON\nExport Bundle"]
Full diagram with DB schema: docs/diagrams/architecture.mmd
| Layer | Technology |
|---|---|
| Framework | Next.js 16 · React 19 · TypeScript (strict + noUncheckedIndexedAccess) |
| Styling | Tailwind v4 · shadcn/ui (new-york) · always-dark design system |
| AI generation | Claude API via Vercel AI Gateway (gateway('anthropic/claude-opus-4-7')) |
| Translation | DeepL API + Claude transcreation layer |
| Orchestration | Inngest (event fan-out, automatic retries, DLQ) |
| Database | Supabase (PostgreSQL + Row Level Security) |
| Eval | LLM-as-Judge harness (AI SDK generateObject with Zod schema) |
| Charts | Recharts (RadarChart · BarChart) |
| Deployment | Vercel (Fluid Compute) |
git clone https://github.com/RexOwenDev/content-factory
cd content-factory
pnpm install
cp .env.example .env.local # fill Supabase + Inngest keys only
pnpm db:migrate # apply Supabase migrations
pnpm db:seed # seed 3 demo projects
pnpm devOpen http://localhost:3000.
All Claude, DeepL, and eval calls run from fixtures/ — no AI API keys needed to explore the full dashboard and review queue.
Set these in .env.local to switch from fixtures to real API calls:
ADAPTER_MODE=live
# AI generation
ANTHROPIC_API_KEY=sk-ant-...
# or via Vercel AI Gateway (recommended on Vercel):
# VERCEL_AI_GATEWAY_URL and VERCEL_AI_GATEWAY_TOKEN set automatically
# Translation
DEEPL_API_KEY=...
# Orchestration
INNGEST_SIGNING_KEY=signkey-prod-...
INNGEST_EVENT_KEY=...See SETUP.md for the full setup guide including Supabase schema, webhook configuration, and Inngest registration.
The seed script populates three showcase projects:
| Project | Markets | Content types |
|---|---|---|
| ForgeTorque Pro 3000 (power tools) | 🇺🇸 🇬🇧 🇫🇷 🇩🇪 🇯🇵 | Product desc · Ad copy · Meta tags · Landing page |
| LuxDermis Advanced Serum (beauty) | 🇺🇸 🇫🇷 🇩🇪 🇯🇵 | Product desc · Ad copy · Meta tags |
| VeloCargo Urban Bike (e-commerce) | 🇺🇸 🇫🇷 🇩🇪 🇳🇱 | Product desc · Ad copy · Landing page |
pnpm db:seed # creates all three with outputs and eval scores pre-populatedsrc/
app/
(dashboard)/ # Projects · Project detail · Review queue · Eval · Analytics
api/
pipeline/generate/ # POST — fires Inngest event
projects/[id]/ # outputs · eval-summary · reviews · export
reviews/[reviewId]/ # PATCH approve/reject
inngest/ # Inngest serve() handler
components/
dashboard/ # ProjectCard · ProjectTabs · StatsRow
pipeline/ # RunPipelineButton · PipelineStatus
output/ # OutputViewer · ExportButton
eval/ # EvalAnalytics (Recharts)
review/ # ReviewCard (HITL)
lib/
adapters/ # claude · deepl · shopify · wordpress (fixture + live)
pipeline/ # prompts.ts
inngest/
functions/ # generateContent · translateMarket
db/
queries/ # projects · jobs · outputs · reviews
fixtures/
briefs/ # 3 product briefs
generated/ # EN-US source content
transcreated/ # Locale-specific transcreations
evaluated/ # Eval score fixtures
shopify/ wordpress/ # Platform output shapes
docs/
diagrams/architecture.mmd
PROJECT-MEMORY.md
scripts/
seed-demo.ts
| Phase | Status |
|---|---|
| 0 — Foundation & Scaffolding | ✅ |
| 1 — Brief Ingestion & Market Matrix | ✅ |
| 2 — AI Content Generation Core | ✅ |
| 3 — Translation & Transcreation | ✅ |
| 4 — Quality & Eval Harness | ✅ |
| 5 — Review Queue & HITL | ✅ |
| 6 — Output Shaping & Integration | ✅ |
| 7 — Dashboard, Analytics & Polish | ✅ |
| 8 — Docs, Visuals & Launch | ✅ |
MIT — see LICENSE