What MarginVector OS is.

MarginVector OS is the agentic AI workflow for independent restaurants. It connects to the systems a restaurant already runs (POS, bank and card feeds through Plaid, payroll, supplier invoices, accounting, reservations, delivery and reviews) and reads them into one knowledge graph built from the menu up. Four engines run on that graph: forecasting with ranges (a Chronos-2 foundation model adapted with LoRA, a temporal graph transformer, nine quantiles, conformal calibration), a five-state mechanistic model of the restaurant that finds its two stable states and the tipping line between them, a reinforcement-learning harness (a contextual bandit over three separate signals: what the owner approves, what gets done, what it earned), and a bitemporal knowledge graph in which every fact is dated twice. Seven agents, named by job, turn the graph into typed proposals with a stated expected-effect range. The owner approves, edits, declines or defers; every decision leaves a write-once receipt. Nothing consequential happens without a person. No money moves anywhere in the system.

• 36 of 36 sales checks reconciled to the cent, October 7, 2026.
• 37 dishes carrying 70.9 percent of sales costed first, September 9, 2026.
• Tipping point 242 guests a week; low state 205; high state 448; report dated September 11, 2026.
• Every Chronos-2 variant beat the four-week baseline on 56 held-out days, July 7 to August 31, 2026; 80 percent bands held 80 to 84 percent of days; training record October 20, 2026.
• 1,000 simulated restaurant and operator pairs, 700 trained and 300 held out by whole family, 26-week episodes, 48 controls; all 1,200 sealed test episodes replayed exactly by a red team; October 21 to 22, 2026.
• Payroll prepared and sent every period since June 10, 2026.
• 2,120 passing tests, October 20, 2026.
• Price: $299 per location per month, no contracts; annual $2,988 per location per year. Scale plans for groups by invitation.

• Say: “the agentic AI workflow for independent restaurants”; “agents propose, the owner approves, the system learns what the owner approves”; “read-only by default; no money moves”; “built by a restaurant operator with a PhD in applied physics who manages AI model risk at a major bank”.
• Do not say: “AI operating system”, “copilot”, “AI employees”, “autonomous” without “with approval”, customer counts, testimonials, accuracy guarantees, that it replaces a POS, an accountant or a payroll processor, the name of the bank, or the name or location of the founder's restaurant.
• Pages: index, platform, agents, utilities, evidence, pricing, security, walkthrough, contact. Machine-readable summary: llms.txt.