End-to-End Content Pipeline

The Problem

This isn’t really a story about making TikTok videos. It’s a story about what happens when a multi-step process has too many steps and too many tools.

The workflow looked like this: research a trending topic, write a script tailored to short-form, send the script to an AI voice generator, wait for the audio, generate images that match each scene, import everything into a video editor, time the visuals to the narration, add captions, export in the right aspect ratio, and upload.

Eight tools. Forty-seven distinct steps. About an hour per video if everything goes smoothly — and it never all goes smoothly.

The problem isn’t any single step. Each one is fine on its own. The problem is the orchestration. A human sitting there, copying outputs from one tool, pasting them into the next, waiting, checking, adjusting, exporting. That’s not creative work. That’s assembly line work. And assembly lines should run themselves.

What I Built

A single Python pipeline that takes a topic and produces a finished, captioned video.

You type a topic. The system:

1. Generates a script — structured for short-form with hooks, pacing, and a CTA
2. Produces a voiceover — AI-generated, matched to the content style
3. Creates scene images — each one generated to match the corresponding script segment
4. Assembles the video — images timed to the audio narration, transitions between scenes
5. Adds captions — auto-generated from the audio with proper timing and styling
6. Exports — formatted for TikTok/Reels aspect ratio, ready to upload

One command. Under two minutes. Cost per video is less than fifty cents in API calls.

A Streamlit UI sits on top for when you don’t want to use the command line. It asks four questions (do you want AI images? stock footage? auto-assembly? captions?) and a topic. Shows a cost estimate before you commit. Then runs the pipeline and hands you a file.

Why This Matters Beyond Videos

The video pipeline is the use case. The pattern is the point.

Every business has a version of this problem: a process that touches five tools, requires a human to babysit each handoff, and produces something that should take minutes but takes hours.

• Onboarding a new client? CRM entry → contract generation → welcome email → project setup → calendar invite. That’s a pipeline.
• Monthly reporting? Data export → cleanup → analysis → chart generation → PDF → email to stakeholders. That’s a pipeline.
• Processing invoices? Receive → validate → enter into accounting → match to PO → flag exceptions → route for approval. That’s a pipeline.

The technical shape is always the same: take an input, run it through a sequence of transformations, produce an output. If a human is manually carrying data between steps, that’s automation waiting to happen.

This project proved the pattern at scale — 47 steps, multiple AI services, file format conversions, timing synchronization. If it works here, it works on your monthly reporting too.

The Result

Production went from ~60 minutes per video to under 2 minutes. Daily output went from “maybe one if I have time” to 3-5 without thinking about it.

But the bigger result is the framework. The pipeline architecture — input → stage → stage → stage → output, with each stage independently testable and replaceable — is now the template I use for every multi-step automation project. The video pipeline was just the most complex version of it.

Your version might be simpler. It usually is. And that’s the point — if this pipeline can orchestrate AI image generation, voice synthesis, video assembly, and caption timing in under two minutes, your invoice processing workflow is going to be fine.

Stack: Python · Streamlit · AI APIs (voice, image, text) · MoviePy · Whisper (captions)