100M Views, Zero Humans: Inside Momentra

Momentra is our autonomous content engine. It detects trends, writes, edits, publishes, and optimises short-form video across platforms without a person in the loop. In its first 90 days it produced over 100 million views.

This is how it works and where it nearly did not.

The loop

The whole system is one closed loop, run continuously:

1. Detect. Monitor signals for emerging audio, formats, and topics with rising velocity.
2. Generate. Produce a script and a video against the detected trend.
3. Score. Evaluate the candidate before it ships. Most candidates do not ship.
4. Publish. Distribute to each platform within its optimal window.
5. Measure. Collect performance per variant.
6. Adjust. Feed results back into generation and scoring.

Step 3 and step 6 are the entire product. Steps 1, 2, 4 are commodity.

Why the scorer matters most

Our first version did not have one. Generate and publish, let the platform decide. Output volume was high and performance was terrible.

The insight: an autonomous system without an internal quality gate optimises for throughput, not results. It will happily publish a thousand mediocre things.

Adding a scorer that rejects most candidates before publication was the single largest improvement in the system's history. Making fewer, better things beat making more things by a wide margin.

The feedback loop is the moat

Anyone can wire a model to a publishing API. What is hard to copy is a system that has been measuring its own output long enough to know what works for a specific audience on a specific platform.

That data does not transfer. It compounds. It is the reason the system got meaningfully better between month one and month three without any change to the underlying model.

What we got wrong

Optimising for views before optimising for consistency. Early on we chased spikes. A viral outlier teaches you almost nothing because you cannot reproduce it. Median performance is the number that indicates whether the system is actually learning.

Underestimating platform difference. The same content does not travel. Each platform needed its own timing, format, and hook conventions. Treating distribution as one step was wrong.

Trusting the loop too early. We let it run unattended before the scorer was good, and it confidently produced volume that did not perform. Autonomy earned by demonstrated performance, not granted upfront.

Where the humans still are

Not in the loop — around it. People set the objective, define what a good outcome is, adjust the guardrails, and decide when the strategy itself is wrong.

The system is autonomous within a frame. Someone still owns the frame, and that is the correct division of labour rather than a limitation to be engineered away.

Does this generalise?

Partly. The architecture generalises to any domain where output is high-volume, performance is measurable quickly, and a bad output is cheap.

It does not generalise where errors are expensive or feedback is slow. That is not a model capability problem. It is structural, and no amount of scale changes it.

More on Momentra.