one page · forwardable · for Rahul → his father
the sovereign raise
how we fund this, honestly
Sovereign intelligence on consumer hardware — ternary inference, agent memory, and the constellation. This is not a pitch deck. It is the plan, the numbers, and the truth, in one page.
peter lodri · peterlodri-sec · 2026.10.08 · Tokyo
1 · what we are building
An AI stack that runs on hardware you own, not on rented cloud:
- Ternary inference — 1.58-bit models (BitNet b1.58 `{-1,0,+1}`) on a laptop CPU. A real engine (`project-zero`) loads and runs them today; live-and-open.
- Context compression — kompress-ultra: a paper, a model, a dataset, and a worker. Keeps ~99% of critical tokens while cutting ~78% of context.
- Agent memory — knowledge-worker (Rahul's): a local-first, provenance-backed knowledge graph that survives between sessions.
- The constellation — a fleet of working surfaces (garden, music, paper, proposal, todos…), all on owned infrastructure.
Everything above is real and linkable today — not a promise.
2 · why now
- Cost collapse. Frontier capability is arriving on consumer silicon. The edge is being sovereign, not just cheap.
- Sovereignty is the product. Regulated industries (health, finance, defence) cannot send data to a third-party cloud. Owned compute + owned models = the only path.
- Timing. The open-model moment (BitNet, ternary, small strong models) is this year. The window is open now.
3 · the pharma bridge (for Rahul's father)
A pharma business runs on two scarce things: compute for protein/AI work, and trust that its molecular data never leaves the building.
- Open protein models (folding / design / binding) can run on owned GPUs — the same sovereign pattern we already do for language models. No per-token cloud bill; no data egress.
- We can pair a pharma business with the sovereign stack: owned hardware + open protein models + a private, provenance-backed data layer. The compounds and the models stay theirs.
- This is not a slogan — it is a concrete, cheaper, more private alternative to renting cloud AI for discovery work.
If Rahul's father wants, we can stand up a private pilot: one owned box, one open protein model, one dataset that never leaves the room.
4 · how we fundraise (the actual plan)
stage 0 · now → 3 months
proof, not promises
- Materials: one-pager, deck, the paper, the model card, a live demo site, a metrics dashboard.
- Metrics to show: tokens saved, benchmark wins, running costs on owned hardware, users.
- Goal: 3 concrete demos and 1 paid pilot conversation. Cost: ~$0 (sweat equity).
stage 1 · friends & family + strategic angels
$25k – $150k · warm intros only
- Warm intros (people who already trust the work), on a SAFE or revenue-share.
- One message, one deck — no spam (Rahul's rule: one big strategic piece).
- What it buys: the box, the runway, a month of shipping.
stage 2 · pre-seed
$250k – $1M · angels, micro-VCs, AI-infra funds
- Show traction first (stage 0 proxies): benchmarks, a paying pilot, adoption of the open artifacts.
- Target funds with a sovereign / edge AI / infra thesis.
- Use of funds: compute, one hire, 12 months of focused build.
stage 3 · seed + strategic
health / pharma strategic
- A strategic investor from the buyer's side (pharma, health AI) contracts before it funds — a pilot that becomes an investment.
- This is where the pharma bridge becomes money, not just a demo.
the non-dilutive lane
grants & community
- Open-source / research grants for the sovereign stack.
- VAKED — a fair-launch, proof-of-work token (21,000,000 cap, no premine, no team allocation). A community/utility funding lane, not a security and not sold as investment. Used carefully, if at all.
5 · the materials checklist
| piece | status |
|---|
| paper (asymmetric-loss compression) | done |
| model + dataset (kompress-v8 / ultrawhale-dogfood) | done |
| engine (project-zero, ternary loader) | done |
| this page (strategy) | done |
| one-pager (PDF) | next |
| investor deck | next |
| metrics dashboard | next |
| pharma pilot proposal | next |
6 · the 90-day plan
| when | what |
|---|
| wk 1–2 | One-pager + deck; the metrics dashboard live. |
| wk 3–6 | 3 demos (compression, ternary engine, agent memory); record them. |
| wk 4–8 | Warm-intro plan: 20 names, one message each, one deck. |
| wk 6–10 | Pharma pilot proposal; 1 conversation with a real buyer. |
| wk 8–12 | Close stage-1 cheques; decide pre-seed timing. |
7 · the honest part
This is a strategy, not an offer. Nothing here is investment advice, a security, or a promise of return. The base rate for any venture is zero — plan for that. If a cheque would hurt to lose, don't write it. We raise on evidence, and we say what we don't know.
And one thing that is already true, no money attached:
∅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 rest passes away. The thread doesn't. — peter <3