Who it is for
Petrichor is built for AI labs and for the models they are training to do science. A customer might be a research team at a frontier lab, a company whose models design experiments it has no bench to run, or an agent working through a problem on its own. What they share is the need to test an idea in the physical world and get back everything that happened while it was tested.
How an experiment runs
From the customer’s side, an experiment moves through six stages.
- The model describes the experiment. It can write in plain language or in a structured job format, and the only limit is what a trained scientist can actually do at a bench.
- We confirm what will be run. A request is checked for instructions that contradict each other. Practical problems on our side, such as a missing consumable, are ours to solve rather than the customer’s.
- Scientists carry out the work. Many experiments run in parallel across the lab, each one handled by trained people.
- Every step is recorded as it happens. Each step is photographed, raw files are exported straight from the instruments, and every action goes into a timestamped log.
- Results stream back during the run. The model sees readouts as they arrive and can change the plan before the experiment is over.
- The complete record is delivered. When the run ends, the customer receives everything it produced, including any steps that failed.
What comes back
Most labs return a final result, whereas Petrichor returns the full record of the experiment, because a model learns from how a run unfolded as much as from how it ended. A record usually contains the following.
- Images
- Photographs of every plate and sample before and after each step.
- Instrument data
- Raw files exported directly from the instruments, in their original formats.
- Step log
- A timestamped record of every action taken during the run.
- Failures
- Steps that went wrong, recorded so the failure is clear in the data.
Speed
Many contract research organisations take around six weeks to return an experiment that takes about five days to run, and most of that time is spent waiting around the experiment rather than on it. Petrichor is organised to remove that waiting, and because results stream back during the run, a model usually starts learning from an experiment well before it finishes.
Questions
What is Petrichor Bio?
Petrichor Bio is a wet lab built for AI. AI models and the teams behind them send experiments to Petrichor, our scientists run them in parallel, and every step is recorded and returned as data.
Can an AI agent submit an experiment directly?
Yes. An experiment can be written by a model in plain language or in a structured format, so an agent can send work and receive results without a person relaying them.
How is Petrichor different from a contract research organisation?
A contract research organisation is organised around human scientists as customers and usually returns a final report weeks later. Petrichor is organised around AI models, runs many experiments in parallel, records every step and streams results back while the experiment is still running.
Does Petrichor train AI models on customer data?
No. Petrichor never trains on customer data, never shares it and never gives it to another lab.
What happens when an experiment fails?
The failure is recorded like any other result, in enough detail to see what went wrong, and it is returned to the customer as part of the data.