private · pilot proposal · not for distribution
sovereign protein AI
open protein models, on hardware you own
A small, scoped pilot: run open protein models — folding, binding, design — on a machine in your building, on your data, with no per-token bill and nothing leaving the room.
prepared by peter lodri · peterlodri-sec · 2026.10.08 · for a private pharma pilot
the offer, in one line
Your molecules never leave your building, your compute is yours, and your AI bill stops being a meter.
the problem with cloud AI in pharma
- Data egress. Sending structures, targets, or sequences to a third-party cloud is a legal and trust wall for a pharma business — often a hard "no".
- Per-token economics. Discovery work is compute-heavy and bursty. Cloud billing punishes exactly the iteration you need to do.
- Lock-in. When the model and the weights aren't yours, neither is the pipeline.
what the pilot is (concretely scoped)
| item | detail |
|---|
| scope | one owned GPU box · one open protein model (folding / binding / design, chosen for your target) · one dataset of yours |
| duration | 4–6 weeks |
| you get | a working private protein-AI endpoint inside your network; benchmarks on your targets; a short honest report; the box is yours to keep |
| you provide | a machine (or we spec one, ~$3–5k capital, one-time) · one dataset · one target to aim at |
| cost model | a modest fixed pilot fee (or paid-in-kind); the box is capital, not rent — no per-token bill |
| data handling | everything stays on the box. We can work air-gapped. Your IP stays your IP. |
why us
- We already run 1.58-bit state-of-the-art open models on consumer hardware — the same sovereignty pattern, applied to protein models.
- We build kompress-ultra (context compression), knowledge-worker (private provenance memory), and the engine (project-zero) — a full sovereign stack, not a wrapper on someone else's cloud.
- We're open, verifiable, and allergic to lock-in: the models are open, the box is yours, the pipeline is portable. If you walk away, you keep everything.
what we honestly do and don't know
Open protein models are strong and improving, but they are not a replacement for lab validation, and results on any given target are not guaranteed. This pilot measures whether sovereign, owned compute meaningfully helps your work — cheaply and privately. If it doesn't, you'll know in six weeks and you'll still own the box. No promises beyond: it stays private, it's yours, and we'll tell you the truth about the results.
the next step
- A 30-minute call — what are you working on, what would be a useful target?
- We spec the box and pick the model together.
- We stand it up behind your firewall and run the first benchmark.
That's it. Small, private, reversible. The worst case is a box you own and a report you keep.
∅R ⟶ᴸ R₁ ⟶ᴸ R₂ ⟶ᴸ ⋯ ⟶ᴸ Rₙ ⟶ᴸ ⌂
│ ρ₀ ⊃ ρ₁ ⊃ ⋯ ⊃ ρₙ ↓ 0
│ π₀ ↝ π₁ ↝ ⋯ ↝ πₙ ↝ π⌂
│ μ(Rᵢ, you) ≠ 0 ∀i ∴ μ(⌂, you) ≠ 0
The compute can be replaced. The thread can't. — peter <3