Petrichor is a wet lab for frontier AI labs and for the models those labs are training to do biology. A model or a research team submits an experiment in plain language or as a structured job, including through a Python library, and scientists on the team run it in parallel with other work. Each step is photographed, raw files are exported from the instruments, and a timestamped log streams back while the run is still going, including failures and steps with no clear result. The customer owns that record. Petrichor does not share it or train on it. The launch comes as AI labs move into biology, where a model cannot check its own answer the way it can in code or math, and the limit is how many real experiments get run. Existing contract labs are built for human scientists and often return a final report weeks later. Petrichor says work that takes about five days at the bench commonly comes back from an outside lab in about six weeks, and it is organized to remove that waiting so a model can see readouts and change the plan before the experiment ends. People still perform every experiment, for safety, until the team is confident robots can take over. Petrichor is in Y Combinator's Fall 2026 batch, based in San Francisco. CEO Wilhelm Hedenskog, previously a YC Fellow, is building it with cofounders Elias Reinfeldt and Jesper Johansson. The three are 19, went to high school together in Gothenburg, Sweden, and moved to the Bay Area for the batch.