The Footnote

The pipeline

1. Find and scrape
   Scrapes candidate stories from the open web and scores them. A model gate rejects weak ones outright.
2. Research
   Builds a sourced brief so the script has facts, not vibes.
3. Write
   A language model drafts the script and repairs it until it validates. A local model can do this job for free.
4. Voice
   Narration through Chatterbox and Gemini TTS. faster-whisper then aligns every word so the captions land exactly on the voice.
5. Illustrate
   Finds real archival images through a self-hosted SearXNG metasearch over museum collections and image indexes. Nothing is generated. Qwen3-VL, a vision model, judges every image against the script before it can appear.
6. Download
   The hard part. The downloader impersonates a real browser with curl_cffi, walks past bot detection on Yandex and Bing, ranks sources by trust, retries through proxies when blocked, and caches everything. Free web-scale image sourcing with zero API bills.
7. Render and publish
   Remotion composes captions, imagery, and audio into vertical video. It publishes twice a day to Instagram, YouTube, and Facebook, with receipt ledgers so nothing ever posts twice.

Design choices that matter

Images are searched, not generated. For history content a real 1943 photograph beats a plausible fake every time. The pipeline runs its own metasearch across museum archives and image indexes, then does the part everyone else pays an API for. It impersonates a real browser at the TLS level, rotates fingerprints and proxies, backs off when a captcha appears, and gets the full-resolution file instead of a thumbnail. That is how the image budget stays at zero.

Why it exists. The whole thing is a free engine for branding and organic traffic. No ad spend, no licensing bills, no creative team, and it compounds while I sleep.

Models are chosen per task, by cost. A strong model writes scripts. Cheaper models gate stories and judge images. The expensive call happens once per video and the cheap calls happen dozens of times, which is why the whole thing runs on almost nothing.

Everything is resumable. Each stage writes its state before moving on. A crashed run continues where it stopped, and a published video can never be published twice.

The stack

Python end to end, structured as seven resumable stages over SQLite state. Language models are assigned by task and price: a frontier model writes scripts, cheaper models gate stories, judge images, and pick thumbnails. Narration is Kokoro TTS running locally, aligned word-by-word with faster-whisper for captions. Images come from a self-hosted SearXNG metasearch over museum collections and image indexes, ranked by source trust. Remotion (React) renders the vertical video. Publishing goes through the YouTube Data API and Instagram on a launchd schedule, with receipt ledgers so nothing ever posts twice.

Same machinery, more channels

The pipeline is multi-tenant. It also runs Iris's entire marketing: AI-generated infographic posters, quality-gated against a written creative contract, published daily as Instagram Reels, Stories, and Pinterest pins. And it fed the animated horror pipeline, which pushes the same idea from slideshows into fully generated film.

Watch

Instagram: @thefootnotemedia · YouTube: @TheFootnoteMedia · Facebook