§ 02 · Products · Press

We replaced four subscriptions with a system.

Press is the media production system Perpetual Core built and operates for its own content — the HeyGen, ElevenLabs, Opus Clips, and Descript jobs running as open-source models on hardware we own. It runs long-form video into scored, captioned shorts, clones a voice from a 3-second sample, and edits from the transcript instead of a timeline.

$0/month to run. The same engineering we apply for clients who want the same system installed under their own brand.

§ 01

What it replaces

Four vendor subscriptions collapse into one system.

01

HeyGen

Talking-head + avatar video

SadTalker, on our hardware

02

ElevenLabs

Voice cloning + narration

Kokoro / Qwen3-TTS, 3-second sample

03

Opus Clips

Long video → scored shorts

Whisper + AI clip scoring, in-house

04

Descript

Transcript-driven editing

Click-to-cut, silence removal, auto chapters

§ 02

Output

Two clips, same source, same system.

Face-tracked framing, burned captions, AI-scored clip selection. The caption style changes per brand voice — the pipeline underneath does not.

Caption style — punchy

Caption style — serif

§ 03

How it works

Transcript in, distributed content out.

01

Ingest.

Long-form video or audio comes in — a keynote, a podcast, a course recording. Whisper transcribes it in roughly 17 seconds per 3 minutes of source.

02

Cut.

The transcript becomes the editing surface: click sentences to cut, silence and filler removed automatically — 22 seconds of dead air stripped from a 3-minute keynote without a human touching a timeline.

03

Score and frame.

Candidate clips are AI-scored for shareability, then face-tracked and framed into captioned 9:16 shorts.

04

Voice and distribute.

Voice-cloned narration batches unattended across long runs — a 63-script course narrated in one pass — and per-brand caption and ad packs go out across up to seven brand voices.

§ 04

The engineering

Open models. Our hardware. No per-seat rent.

The stack is Apache- and MIT-licensed open models — Whisper, Kokoro, Qwen3-TTS, SadTalker — running on Apple Silicon we own, orchestrated by Claude agents. No HeyGen render queue, no ElevenLabs character limit, no per-seat Descript license.

A performance loop closes the pipeline: real analytics CSVs feed back into clip selection, so the system gets better at picking winners instead of running the same scoring heuristic forever.

Nothing here is a demo. It runs our own content pipeline today.

§

By invitation

Press is not self-serve. It is a system we build for operators.

We install the same pipeline under your brand voice, on your archive, pointed at your distribution. Tell us the archive you have and the channels you publish to.