Automated Content & AI Agent System

Producing content for six TikTok channels was not mainly a writing problem. The real cost lived in the handoffs: finding a topic, writing a script, generating visual material, preparing voice and subtitles, editing vertical video and recovering when one service failed.

Building a production line, not a prompt

The system uses n8n as the coordinator and keeps generation services replaceable. LocalAI, ComfyUI and Qdrant run beside the workflow in Docker, while a dedicated FFmpeg service handles the part that language models cannot: turning assets into publishable video.

That media service accepts a task rather than a loose shell command. It can create Ken Burns sequences from still images, split long source footage, crop a selected segment to 1080×1920, apply overlays and colour treatment, add subtitles, mix voice with background music and return a predictable download URL.

Why the infrastructure matters

The stack was prepared for an Ubuntu host with an RTX 3090 and CUDA 12.2. GPU, automation and local-model workloads can be started as separate Compose profiles, so an unavailable image service does not require rebuilding the whole system.

The result is a pipeline whose stages can be observed and retried. That is what makes repeated content production economical: not that generation is free, but that the same operational path can run again without rebuilding every handoff by hand.