[work] ai newsroom
An autonomous newsroom.
Public TV's pipeline was not broken. It was built, professionally, for the 8pm bulletin, while the audience moved to feeds that never sleep. Their question: what can AI actually do here? Our answer: an autonomous newsroom, from live feeds to published clips, with one operator supervising.
[01] the bottleneck
Five to eight people in sequence. 45 to 120 minutes per clip. Most of it assembly, not journalism.
A media company with a decades-deep broadcast operation came to us with a sharp brief: how do we switch up the social media game, and what can AI actually do here? Nothing about their pipeline was broken; it was built, professionally, for a different world. The 8pm bulletin, the studio floor, the broadcast tower.
We started where we always start: how the work actually happens. A short-form clip touched editor, researcher, writer, producer, voice, video editor, motion designer, and reviewer in sequence. The stories aged faster than the pipeline could package them. That is not a talent problem. It is a shape problem.
[02] editorial agents
Ingest -> Enrich -> Curate -> Script -> Voice -> Presenter -> Composite -> Review -> Package.
The Hot Take loop: a Director agent discovers and curates stories near-instantly from live feeds. Parallel Researcher agents deep-dive the strongest candidates. A Writer turns research into script, a Stage Director adds shots and pacing, an Editor reviews, and a Sponsor agent reads the finished piece and places contextual ad segments at natural breaks. Mission Control shows an operator what every agent is doing.
[03] presenter layer
Designed voices. Lipsynced presenters. Open-source models, self-hosted.
Ultrarealistic talking heads at channel-grade quality, running on open-source models we self-host on RunPod servers spun up for production. That is what makes the economics work at daily volume, where hosted alternatives price a single 30-second clip in the multi-dollar range.
When the format demands cinema, the same editorial brain drives an Unreal Engine pipeline: MetaHuman performance, automated sequencing, film-grade rendered output.
[04] the cockpit
Every stage observable, editable, and re-renderable.
An operator cockpit where every stage can be inspected and re-rendered without restarting the job. Branded compositing so the output looks like it belongs to the channel, not to a template. Editorial memory underneath: beats tracked across days, story packages reused across shorts, bulletins, deep dives, and sponsor inventory.
[05] the outcome
One operator supervising instead of five to eight in sequence. Minutes instead of hours.
An autonomous newsroom: catching beats as they break and publishing to Instagram, YouTube, and wherever the audience actually lives, with a minimal human in the loop. The human does not disappear; the human moves from assembly to judgment, sourcing, taste, and the publish decision. The economics are stated as the operating model the system is built to, not audited telemetry.
The machine generalizes. Sports highlights, market commentary, regional-language news, branded explainers: any operation where stories arrive faster than humans can package them, and the package has to look like it belongs on a specific channel.