← All work

AI content automation platform

AI automation · RAG · content pipeline · Theology Academy, MENA

Topic in, finished video out. An Airtable-driven n8n pipeline sitting on top of a retrieval system that grew to roughly 71,000 vectors — built deliberately with human approval gates rather than full autopilot, because content nobody checked is content nobody can publish.

The system, running

A walkthrough of the pipeline that turns a source library into finished YouTube video — Airtable control surface, n8n workflow, approval gates.

The work

Problem
Take a topic and produce finished video content end to end, at volume, without a person hand-driving each step — but also without letting unreviewed AI output reach an audience.
Control surface
Airtable as the operator's console — Topic, Scripts & Voice-Over, Library and Variables tables — driving the whole n8n pipeline. Non-technical operators run it from a spreadsheet view rather than a workflow editor.
RAG system
Book / PDF → extraction → chunking → embeddings → vector DB → retrieval → LLM generation
Grew to ~71,000 vector rows at 1536 dimensions across Supabase/pgvector, Qdrant and PostgreSQL, with source material and rendered assets on DigitalOcean Spaces and GCS.
Prompt rotation
Around 20 prompts tracked with active status, sequence position, last run and run count. The pipeline selects the next prompt itself instead of waiting on a human to remember which one came last — the difference between a system that runs and a system someone has to babysit.
Approval gates
Staged: Create Script → Generate VO → Generate Shots → Create Final Video. Each stage is reviewed before the next one spends tokens or render time. A bad script gets caught before it becomes a bad video.
What broke
Token limits, chunk sizing, Unicode errors in source texts, Supabase timeouts on large queries, large-dataset handling, and PDF extraction failures. Most of the engineering was here — at 71,000 vectors the naive version of every step falls over.
Stack
  • n8n
  • Airtable
  • Supabase
  • pgvector
  • Qdrant
  • PostgreSQL
  • OpenAI embeddings
  • DigitalOcean Spaces
  • GCS
  • Docker
Outcome
A running content pipeline an operator drives from Airtable, producing finished video from a source library with review points where they matter.

What it produces

Generated scene — candlelit assembly
Generated scene art, retrieved and prompted from the source library
Generated portrait illustration
Illustrated portrait, held to a consistent house style
Generated title card
Title card — the pipeline assembles the shot, not just the image

This is one of a few projects shown here — a small sample of what I've built. Happy to go deeper on the retrieval architecture or the approval design.

Contact

Currently exploring AI filmmaking and automation roles. Happy to walk through any of these systems in detail — architecture, trade-offs, and what broke along the way.

Résumé (PDF) ← All work