project25.live BACK

A closed loop, no human in the path.

Project25 is one system, not a tool. It writes content, publishes it from a fleet of ~1,968 real devices, measures exactly what earned attention, and rewrites the next batch, continuously and autonomously. This is the high-level shape of how it works, and the parts that are genuinely hard. The deep version, we save for people who build it with us.

THE PIPELINE

Five stages, one loop.

01
Generate

Thousands of content variants per cycle (hooks, edits, formats) produced from the current model of what earns attention. The interesting part isn't generation; it's that nothing here is hand-picked.

02
Review

Every variant is scored before it ships: quality, policy risk, and a prediction of how it will perform. The model decides what's worth a slot on a real device. No human approves a post.

03
Publish

Winners go out from a fleet of ~1,968 real devices. Each device has to look and behave like a real person to platforms that are actively trying to detect it. That is the hardest sustained problem we run.

04
Measure

Every impression, every retention curve, every drop-off is captured and attributed back to the exact variant that produced it. The loop is only as good as this attribution.

05
Learn

Results feed straight back into the model that drives Generate and Review. The next cycle is measurably better than the last, with no one in the middle.

THE HARD PARTS

Where the real work is.

Most of the system is plumbing. These four are the problems we actually lose sleep over.

Survival

Keep a fleet of real devices alive against platforms that fingerprint, behaviour-model, and ban in waves. Every signal (network, motion, cadence, identity) has to be coherent and human, continuously, at fleet scale. This is the problem that never stops.

Closing the loop

Run generate → measure → learn with no human in the publish path. That means trustworthy automated review, safe-by-construction publishing, and a measurement layer good enough that the model can be left alone with it.

Attention priors

Predict which variant will earn attention before a single view is spent, so the fleet's limited, expensive slots go to the bets most likely to compound.

Self-recovery

Detect when a persona drifts or an account degrades and recover it autonomously (re-warm, re-route, repair) without a human noticing it happened.

HOW WE WORK

Four people, 1.2B views.

Small on purpose

Every engineer owns a layer of the loop end to end. The team stays small so ownership stays whole.

Ship daily

No managers, no busywork. The system is live; changes land against real devices and real attention, today.

Adversarial & real

This is not a clean research sandbox. The platforms push back; the work is messy, real-world systems engineering with a live adversary.

Want the deep version?

The real architecture (the survival stack, the measurement layer, the recovery loop) we walk through with engineers we’re hiring. Tell us which hard part you’d take.