PX-ReelFlow — AI Content Pipeline

Use case

A fully automated pipeline that turns a topic hint into a complete short-form clip: an LLM produces the quote, an image model renders the background, a video model animates it, a music model adds a soundtrack, FFmpeg composes everything into a finished reel, and a worker posts via the Instagram Graph API. Human-in-the-loop approval gates through a React dashboard.

What's done so far

• Backend: FastAPI with async SQLAlchemy and asyncpg, APScheduler for recurring jobs, Telegram bot for mobile control
• Frontend: React 19 + TypeScript + Vite 7 + Tailwind v4 + Zustand for state management
• Persistence: PostgreSQL 16 as job queue with FOR UPDATE SKIP LOCKED — no Redis needed
• Worker: Python poll-worker on a GPU machine pulls jobs, calls ComfyUI for image/video, HeartMuLa for music, FFmpeg for compositing
• AI stack: Z-Image Turbo (image), LTX-2.3 22B FP8 (video), HeartMuLa 3B (music)
• LLM layer: switchable between local Ollama models, Claude Bridge, OpenAI, Anthropic, and OpenRouter
• Instagram integration via the official Graph API with nginx-hosted public URL for video upload

In progress

Hardening the approval gates, more robust retry behavior in the pipeline, per-run cost tracking, multi-channel support. The project sits inside the Pradox AI Incubator and is primarily an end-to-end lab for AI content tooling — no commercial monetization.