Si-Mind is a sovereign decision substrate. Capture your life — voice, photo, text — through infrastructure you own. Ask hard questions of it. Get decision-grade answers synthesized adversarially across multiple models, with provenance back to the moment you captured it.
Nobody else can do this. Because nobody else has your data. And we don't keep it either.
Capture lands in your private graph. The synthesis surface reads from the graph and reasons over it across multiple models simultaneously, each citing back to the captures that informed it. The decision goes back into the graph. The next question is smarter — about you, with you.
ChatGPT can't do this; it has no graph of your life. Anytype can't do this; it has no adversarial synthesis. We have both. The switching cost rises with every captured thought.
One substrate, multiple altitudes. Self-host or use a managed surface; the interface is identical — only the chrome and what's editable differ.
These aren't aspirational claims. They're constraints built into the architecture from day one — the substrate cannot do otherwise without violating its own canon.
Your data lives on infrastructure you control. The company doesn't see your captures. Self-host or use a hosted instance — your call, never ours.
Every claim Si-Mind makes traces back to the capture that produced it. No floating context. No "the AI said so." Receipts always.
The right models, orchestrated across providers, argued against each other when the question is hard enough to warrant it. The synthesis is what survives the disagreement. Single-model confidence is the trap.
You're weighing a strategic call — say, whether to take a partnership offer that would change your runway. You ask Si-Mind. It pulls forty-three captures from the last six months: a voice memo from a tense call with the would-be partner, financial notes from your Q1 review, a journal entry about why you started the company.
A panel of frontier models reads the graph in parallel — pulled per-question from Claude, Grok, Gemini, Perplexity, GPT, and others as they evolve. Claude flags hesitation in your own captured tone two weeks ago and recommends pausing. Grok surfaces a numerical inconsistency between the partner's projection and your own gross-margin math, citing the exact capture. Gemini pulls a precedent from your own past decisions that argues against the deal.
The synthesis is not a vote. It is what survives the disagreement — the structurally durable claim once each model's case has been pressure-tested by the others. Every line carries a footnote back to the capture that informed it. You can audit the reasoning at the same depth you'd audit a financial statement.
Retrieval quality is measured weekly and published. If it doesn't improve, the work that week reverts. No vibes. No marketing language for capabilities we can't measure.
The substrate is AGPL — copyable, inspectable, forkable. Commercial license for proprietary use. No hostage software, no surprise paywalls.
Twenty-plus questions drawn from the founder's actual work — factual lookup, synthesis, cross-domain. A separate judge model scores each answer 0–5 on faithfulness to the cited sources. The full suite runs weekly against the live substrate. Every change to chunking, embeddings, retrieval, or context assembly is A/B'd against the current production baseline before it goes in.
If the score trends down after a change, the change gets rolled back. The work for that week reverts to a shadow branch until we understand what regressed. No silent quality drift.
This is the practice we ship — refined against our own slop, not benchmarked into a press release.
v0 is in build. Seventeen-week plan, shipping end of September 2026. iOS Companion + Si-Tunnel transport + Si-Brain substrate + Si-Prism synthesis, fused into one product. AGPL release at launch. Private TestFlight beta from week 13. The eval suite, judge prompts, and dataset all ship with the code.
Built by Tony Proia. A year of running multi-agent systems on local infrastructure on Apple Silicon. The substrate I needed to think clearly, built in the open.
No marketing — just a note when the AGPL release lands and TestFlight invitations open. Maybe one more if there's something genuinely worth telling you.