PERPETUAL LABS
The data & deployment layer for AI.
Every business has two parts
The first is the relationship: the part that matters most, and the part that must stay human.1 The second is everything around it: the scheduling, the claims, the dispatch, the documentation, the billing.2 The machine. It decides whether the relationship gets your best hour or your most tired one.
AI should run the machine. Today it can't: how to run a business was never written on the internet. It lives in systems of record, in customer conversations, in the heads of experienced operators. Nobody ever scraped it, labeled it, or published it. And it shifts with every account, every regulation, every handoff.
1 The doctor with the patient. The coworkers at a whiteboard. The accountant a business owner trusts.
2 The machine, enumerated: scheduling, claims, dispatch, procurement, documentation, compliance, planning, billing, follow-up.
"Models learned from the internet. That's why they're superhuman at code, math, and text, and still can't run a lemonade stand."
Perpetual closes that gap.
We turn how work is done into intelligence that does the work.
The Foundry
We turn real work into training. The traces, workflows, and judgment calls your people make every day become the data that post-trains a model. You choose what to capture.
PerpetualOS
The post-trained model runs your machine, and in your domain it's faster, cheaper, and more accurate than frontier intelligence. Every action it takes writes a new trace.
MODEL ▸
data · workflows · learnings
Figure 1. The machinery, live. The Foundry turns traces into training assets; PerpetualOS runs the resulting model, and every action writes new traces.
Every deployment sharpens the data. Sharper data post-trains a sharper model. Each generation runs more of your machine, handing your people back their time for the one thing only they can do: the relationship.
Figure 2. The model ladder, generation by generation. Hover a bar to see what each version runs on its own.
Who we are
We've done each piece of this work at the places that defined it. Perpetual is those three crafts, run as one loop.
We ran some of the largest data programs the frontier ever bought. Moonshots, coding, expert knowledge work. We learned exactly which data makes a model better.
Then we built the other side: post-training, evals, and RL across coding, science, and audio. We know what actually sharpens a model, and what just burns compute and tokens.
Finally, we've shipped agents into production for F500 enterprises, in operations where a wrong answer costs real money.
Let's build something that runs forever.
For operators of essential businesses, and for teams building models. Tell us what you run.