AI Content Studio
The agent workforce that runs Ballknowers — six specialised agents on a task queue, every publish path gated by human approval
- runs a live consumer app
- human-gated publishing
- voice-operated

Overview
A separate system that operates a live consumer app. Specialised agents — writer, analyst, artist, editor, publisher, support — poll a Supabase task queue, draft articles, social posts and video, and route everything through a human approval gate before a one-way publish into the product. A two-database factory/storefront split keeps drafting isolated from what users see. Operated through a HUD showing every agent's live status, a Content Hub and Writing Room for review, a Remotion render engine for video, a custom MCP server for article CRUD, and a voice interface — wake-word detection, local Whisper transcription, ElevenLabs speech — so the studio can be driven hands-free.
The challenge: A consumer app needs constant content to stay alive, and one person cannot write, illustrate, edit and publish it daily. The studio is the workforce that does it — but nothing reaches users without me approving it first.
Agent Orchestration
Six specialised agents poll a Supabase task queue through a Python worker. Drafts move through quality review and a human approval gate, then publish one-way into the consumer app across a two-database factory/storefront boundary.
Voice Interface
Porcupine wake-word detection, local Whisper speech-to-text, and ElevenLabs speech, driving the operator HUD by natural language so the studio can be run without touching a keyboard.
Programmatic Media
A Remotion render engine produces social video and carousels from structured data, with ComfyUI and Blender workflows feeding the assembly line.