Your science,
in the weights.
A post-training lab for drug makers. We turn your proprietary data into an open model that reasons like your scientists, and runs in your own cloud.
Compound 7b loses tenfold potency against 7a once the methyl moves to the meta position. Give the most likely reason and cite the assay.
Before Phase I comes Phase 0: a microdose to learn fast and cheaply. We work the same way. One workflow, one benchmark, proof before scale.
Phase 0, or exploratory IND, studies use doses too small to treat anything. Their only job is to show early whether a candidate is worth the cost of everything that follows.
Four stages,
your model.
First, ELNs, LIMS and assay archives become a clean training set.
Then your scientists’ judgment becomes a private benchmark.
Then fine-tuning and reinforcement learning on open weights, scored against that benchmark at every checkpoint.
Last, it ships to your cloud. The weights are yours.
Deep data,
scarce experts.
Start a pilot
We come from post-training at OpenAI and Google DeepMind, and from AI-for-science groups. Now we do that work for one industry.
The long arc: from models that read your science, to models that help do it.
Bring one workflow and the data behind it. We return the same model, reshaped, with the benchmark that proves it. Start your Phase Zero.
A pilot is one dataset, one benchmark and one post-trained model, scoped on a first call. Everything we build stays with you.
Pilot enquiries will open soon.