Content Factory — AI-Powered Brand Content Pipeline

Section 2: Architecture

2.1 Infrastructure — Hetzner VPS (Recommended for MVP)

Server: Hetzner CPX41 — 8 shared vCPU, 16GB RAM, 240GB NVMe — ~$21/mo

Docker Compose — 6 containers:

Monthly cost: ~$26 (server $21 + backup $5)

Migration path to Fargate: 5–8 days effort when revenue justifies $300+/mo.

2.2 Alternative Deployments

VPS Provider Comparison (8 vCPU, 16GB)

2.3 Database Architecture

Phase 1: Simple PostgreSQL 16 — plain JSONB for content drafts, brand configs, publish logs, cost ledger. No vector/graph extensions needed.

Phase 2: Unified PostgreSQL 16 with extensions:

Scale-up path: Split to dedicated FalkorDB (graph) + Qdrant (vector) when needed.

Section 4: Video Generation Benchmark

4.1 Benchmark Reference

Benchmarked against the northern_lights_iceland reel produced with neural-reels (Veo 2.0 via Vertex AI):

4.2 Per-Clip Cost Comparison (8s, 720p, 9:16)

4.3 Seedance 2.0 — Status & Proxy Warning

Seedance 2.0 is geo-locked to mainland China (Jimeng/Doubao). International API launch indefinitely delayed due to copyright lawsuits from Disney, Netflix, Paramount, Sony, and Warner Bros.

Proxy services analysis: Most services advertising “Seedance 2.0” actually serve Seedance 1.5 Pro. Only Kie AI credibly had 2.0, but deactivated it.

Recommendation: Do NOT rely on Seedance 2.0 proxies for commercial production. Use Seedance 1.5 Pro (legitimate, stable) or alternatives at similar quality: Veo 3.1 Fast, Kling 3.0, Hailuo 02 Pro. Seedance 2.0 quality is meaningfully better than 1.5 Pro (9/10 vs 7.5/10), but compliance risks (ToS violation, data privacy, service stability) make proxy access unsuitable for production.

4.4 Quality Rankings (Early 2026)

1. Runway Gen-4.5 (Elo 1247) / Veo 3.1 (Elo 1226) / Kling 3.0 (Elo 1225)
2. Seedance 2.0 / Sora 2 (9/10 quality — geo-locked to China, no international API)
3. Hailuo 02 (#2 on Artificial Analysis benchmarks)
4. Seedance 1.5 / Vidu Q3 / Pika 2.2
5. Wan 2.5 (best open-source) / LTX-2.3

4.5 Video Strategy by Tier

4.6 Template-Based vs AI-Generated

85–90% of business reels are template-based, not AI video generation:

Section 6: Phase 2 — Knowledge Base + Intelligence

6.1 Phase 2 Overview

Goal: Add competitor intelligence, centralized knowledge base (RAG), analytics feedback loops, and content performance optimization. Agents now learn from past content, competitors, and engagement data.

What Phase 2 adds:

• Competitor scraping (Instagram, LinkedIn, Facebook, X, TikTok)
• pgvector + Apache AGE extensions for vector search and graph relationships
• RAG service for brand knowledge, competitor patterns, past content performance
• Analytics collection (engagement metrics from all platforms)
• Weekly performance reports via Slack
• SEO pipeline for blog articles
• Content performance feedback loop (what works → more of that)

6.2 Data Flow (Phase 2 additions in bold)

flowchart LR
  subgraph INGEST[" Data Ingestion "]
    direction TB
    IG["Instagram\nScraper"]
    LI["LinkedIn\nScraper"]
    FB["Facebook\nScraper"]
    RSS["RSS / News\nFeeds"]
  end
  subgraph DB[" PostgreSQL "]
    direction TB
    JSONB["JSONB\nDocuments"]
    VEC["pgvector\nEmbeddings"]
    GRAPH["Apache AGE\nGraph"]
  end
  subgraph AGENTS[" AI Agent Team "]
    direction TB
    RESEARCH["Researcher\n(Gemini 2.5 Flash)"]
    WRITER["Writer\n(Claude Sonnet)"]
    EDITOR["Editor\n(Claude Sonnet)"]
    ROUTER["Router\n(Gemini 2.0 Flash)"]
    MEDIA["Media Producer\n(Flux + Seedance)"]
    RESEARCH --> WRITER --> EDITOR
    ROUTER -.-> WRITER
    WRITER -.-> MEDIA
  end
  subgraph HITL[" Approval "]
    direction TB
    N8N["n8n\nWorkflow"]
    SLACK["Slack\nApprove / Reject"]
    N8N --> SLACK
    SLACK -->|Revise| N8N
  end
  subgraph PUBLISH[" Publish "]
    direction TB
    CAL["Postiz\nCalendar"]
    LI_PUB["LinkedIn"]
    IG_PUB["Instagram"]
    FB_PUB["Facebook"]
    X_PUB["X / Twitter"]
    TT_PUB["TikTok"]
    YT_PUB["YouTube"]
    BLOG["Blog"]
    CAL --> LI_PUB & IG_PUB & FB_PUB & X_PUB & TT_PUB & YT_PUB & BLOG
  end
  INGEST --> DB --> AGENTS --> HITL -->|Approved| PUBLISH
  PUBLISH -.->|Analytics| DB

**Scraper (n8n scheduled) → Competitor data → PostgreSQL (JSONB + pgvector)**
RSS/News (n8n scheduled) → Ideation agents → Content ideas ranked
                         → **Knowledge base: brand voice, competitor patterns, past winners**
                         → Agent team → Draft → Approval → Publish
**Analytics (n8n scheduled) → Engagement metrics → Performance feedback → Ideation**
**Weekly Slack report → Next week's ideas + last week's performance**

6.3 Phase 2 Waves

Wave 5 — Competitor Intelligence (units 23-27)

Wave 6 — Knowledge Base + RAG (units 28-31)

Wave 7 — Analytics + Optimization (units 32-36)

Phase 2 Dependency Graph

Phase 1 complete
       ↓
Wave 5 (23-27): [pg-extensions] [ig-scraper] [li-scraper] [fb-scraper] [x-tiktok-scraper]
       ↓
Wave 6 (28-31): [competitor-analysis] [rag-service] [enhanced-researcher] [enhanced-ideation]
       ↓
Wave 7 (32-36): [analytics] [weekly-report] [seo] [feedback-loop] [ab-variants]

Section 8: Brands & Competitors

8.1 Brand Profiles

All competitors have active LinkedIn + Facebook + Instagram presence.

dokwallet (Crypto Wallet)

doktransfers (File Transfer)

vesappsai (AI Chat)

Section 10: Premium Models (March 2026)

10.1 Text Models

10.2 Image Models

10.3 Video Models — Cloud API

10.4 Video Models — Self-Hosted / Open Source

Key findings: RTX 4090 (24GB) can only run CogVideoX-5B, LTX-2.3, and HunyuanVideo v1.5 (with offloading) at 720p. Wan2.1-14B and HunyuanVideo full require A100 80GB or H100. Seaweed-7B (the model behind Seedance) weights are not publicly available — self-hosting is not possible.

GPU Requirements Summary

Self-Hosting Cost Per Video (Best Providers)

Self-Hosting vs API Break-Even

Dedicated Server Option (Hetzner)

Self-hosting economics: For 24 reels/month, API is more cost-effective. Self-hosting breaks even at ~100+ videos/day (LTX-2.3 on cheap GPUs) or when using HunyuanVideo with full optimization stack on H100. Hetzner dedicated is best for continuous workloads — GEX44 at $204/mo can produce ~14,400 videos/month at $0.014 each. Note: Hetzner GPU prices rising ~30-37% from April 2026.

10.5 TTS Models