Fold structures and design binders with open models — Boltz-2 and ESMFold2/ESMC — on a machine in your building. No per-run cloud. Nothing leaves the room.
Protein AI is stuck behind the same wall as everything else: it rents compute from someone else's cloud, and it wants your molecules to come along. For a regulated lab or a pharma business, that's not a pricing problem — it's a trust wall.
This lane is the answer: the same folding and design science, running sovereign.
| the cloud way | the sovereign way | |
|---|---|---|
| compute | rented (Modal, etc.) | owned GPU box |
| your data | leaves your network | stays in the room |
| cost | per run | capital, not rent |
| walk away | you lose the pipeline | you keep everything |
Sequence → structure with Boltz-2 (optional MSA server), viewed in Molstar.
the proven surface ›Live fold/design dashboard (Modal-hosted, 8b.is members): protein-q-fold-bind-ui. The work here is moving that surface onto owned silicon.
Folding is for molecules; registration is for anatomy. Built with standardgalactic (Nate): biomechanical_DLIR — physically-constrained deformable image registration for medical imaging. Same sovereignty question on the clinical side.
One platform: model the structure, own the compute, keep the data.
A private, scoped pilot for a pharma business: one owned GPU box, one open protein model, one dataset that never leaves the room — 4–6 weeks, and the box is yours after.
the constellation · garden.vaked.dev · github.com/8b-is/proteinFolding
for Rahul, my brother — and for his father. <3