The agentic AI workflow for independent restaurants

Your restaurant, run by agents that learn.

Seven agents read the records you already produce, propose priced moves on costs, prep, orders, schedules, cash and pricing, and change nothing until you approve.

  • 96

    utilities

  • 4

    engines

  • 7

    specialist agents

  • 9

    forecast quantiles

  • 2,120

    tests

The workflow

It proposes. You approve. It learns.

Proposal 418 · Menu and pricing agent · Mon 06:40Demo numbers

Reprice two dishes after Sysco invoice 48812

  • Mozzarella$3.40 to $4.20 a pound +24%
  • Margherita$16.00 to $16.50
  • Lasagna$18.00 to $18.50
  • Expected effect+$470 a month
  • Range, 80%+$310 to +$620
  • Cash next month, low sideup $160

Approved Mon 06:42 · receipt 9f3a1c7e

  • Every proposal states its expected effect as a range and cites the invoice, the recipe and the sales it came from.

  • Nothing consequential happens without a person. Approve, edit, decline or defer; every decision leaves a write-once receipt.

  • Reinforcement learning over three separate signals, what you approve, what gets done and what it earned, ranks the next proposal.

  1. 36 of 36 sales checks reconciled to the cent October 7, 2026
  2. 37 dishes, 70.9% of sales costed before any price moved September 9, 2026
  3. 242 guests a week the tipping point, found September 11, 2026
  4. Every period since June 10, 2026 payroll prepared and sent Workspace record
  5. 2,120 passing tests October 20, 2026

The graph

Built from the menu up.

Every fact dated twice

fact
Mozzarella $4.20 a pound
valid from
Oct 19, 2026
known at
Oct 19, 2026 06:10
supersedes
$3.40, known Sep 2, 2026

One graph links your menu to its dishes, recipes, ingredients, invoices, shifts, sales and ledger.

A mozzarella price on one Sysco invoice reaches two recipes, two dishes, their contribution and the cash plan.

Every fact is dated twice: when it was true, and when the system learned it. A correction made today never rewrites last month.

  • 14,527

    nodes

  • 103,783

    edges

  • 47

    node types

  • 50

    edge types

A founding restaurant's graph, reproduced exactly by the parity gate, October 2026.

Forecasting

Forecasts you can plan to.

Demo numbers
ObservedLikely80% range

Daily sales, 28 days observed and 14 ahead. Dollars per day.

A Chronos-2 foundation model adapted to your restaurant with LoRA, a temporal graph transformer that reads your menu structure, and nine quantiles per day turn sales history into a range you can plan prep, orders and staffing to.

80 to 84%

Every Chronos-2 variant beat the four-week baseline on 56 held-out days, July 7 to August 31, 2026, and the 80 percent bands held on 80 to 84 percent of days.

Training record dated October 20, 2026

The restaurant model

A model of your restaurant, not just your sales.

Fixed-cost cut per week $0 a week

Five states, customers, perceived quality, staffing capacity, cash and price, fitted to your records with a particle filter.

It finds the two states a restaurant can live in and the line between them, in guests per week.

It tests a staffing, pricing or fixed-cost move before you make it. Cut fixed cost by $1,037 a week and the low state disappears. Commit before the season, not into the trough.

  • 205

    low state, guests a week

  • 242

    tipping point

  • 448

    high state

From a live restaurant's fitted model, report dated September 11, 2026. The 205, 242 and 448 marks are from the report; the flow between them is the model's reduced form.

Learning

It learns what you approve.

A contextual bandit, the reinforcement-learning method that ranks choices by their outcomes, learns from three separate signals, so a click never counts as profit and the next proposal starts from your priorities.

  • Preference. Explicit approve and decline only. Silence is never rejection.
  • Completion. Authenticated execution receipts only. A plan is not a result.
  • Profit. Realized ledger profit against a frozen baseline, after the declared 4-, 8- or 13-week horizon.

1,000 pairs · 700 trained · 300 held out

Trained on 700 simulated restaurant and operator pairs and scored on 300 held out by whole restaurant family: 26-week episodes, 48 controls, 47,215 weekly decisions in training. Four behaviors compared head to head, frozen RL, a one-week planner, a four-week planner and uniform exploration, and all 1,200 sealed test episodes replayed exactly by a red team.

October 21 to 22, 2026

The whole workflow

Seven agents. One week.

MarginVector OS · Demo restaurantWeek of October 19, 2026 · Demo numbers

Monday 06:10 · Operations and menu agents

Invoices land.

Sysco invoice 48812 read; 14 ingredient prices update across the graph; two reprice proposals arrive with their expected effect.

Approved 06:42Margherita $16.50 · Lasagna $18.50receipt 9f3a1c7e

Tuesday 05:30 · Forecast and inventory agents

The week's forecast and prep plan.

A daily range for the next 14 days; prep at the 80th percentile; an order to the safe side: 6 cases mozzarella, 4 cases tomatoes.

Approved 07:15order draft ready for your signaturereceipt 41be02d9

Wednesday 06:00 · Operations agent

Schedule to the likely side.

Servers per shift at the 50th percentile, one on call at the 80th for Friday and Saturday; labor at 29.4 percent of forecast sales.

Edited 08:02one more server Fridayreceipt c7d0e815

Thursday 09:00 · Advertising agent

Weekend special, scored.

The advertising agent drafts the weekend special; the nine-dimension rubric scores it 9.2 of 10; print checked at 300 dpi.

Approved 09:40you post itreceipt 2a9f6b30

Friday 23:10 · Forecast agent

Service, then the range check.

Friday sales $4,310 against a likely $4,120 and an 80 percent range of $3,480 to $4,760.

Inside the rangeforecast receipt records the hitreceipt 5d2e9c47

Saturday 06:00 · Bookkeeping agent

Cash plan, low side.

The 13-week cash plan is rebuilt on the low side of the range: payroll, rent, Sysco terms. Minimum cash $11,900 in week 6, above the reserve.

Reviewed 07:05no action neededreceipt 88d1f4a2

Sunday 23:30 · Bookkeeping and intelligence agents

Close.

POS to deposits, tips and fees reconciled; the weekly P&L drafted; every approval listed with its receipt; the learner updated from four approvals, one edit and no declines.

Closed 23:48journals to CPA reviewreceipt e3c75d1b

Week of October 19, 2026 · Demo numbers

P&L.

Sales$24,860
Food cost (29.0%)$7,210
Labor (29.4%)$7,310
Other operating$2,140
Fixed cost$5,889
Operating result$2,311

Receipts 9f3a1c7e 41be02d9 c7d0e815 2a9f6b30 5d2e9c47 88d1f4a2 e3c75d1b

Monday. The invoices land and two reprices are proposed, priced and approved by 06:42.

Tuesday. The forecast sets prep and an order to the safe side of the range.

Wednesday. The schedule follows the likely side; you add one server for Friday.

Thursday. The weekend special is drafted, scored 9.2 and approved. You post it.

Friday. Sales land inside the range, and the forecast receipt records the hit.

Saturday. The 13-week cash plan is rebuilt on the low side. Nothing to do.

Sunday. The week closes, the P&L is drafted, and the learner updates from what you approved.

One week, every agent, seven receipts, and a P&L you can open to its sources.

The utilities

96 utilities. Twelve areas. One graph.

P&L and balance sheets Reconciliation and exceptions Balanced journal proposals Close readiness Cash, accrual and operating views Invoice and credit evidence Accounting exports Financial audit trail 13-week cash scenarios Obligations and shortfalls Optional-spend comparison Funding and credit scenarios Purchase cash reservations Operating-change scenarios Recipe and ingredient costing Item contribution Supplier price changes Price-elasticity estimation Price and volume tradeoffs Versioned recipe changes Recipe-based ingredient demand Daily and shift ingredient COGS Menu strategy portfolios Sales trends and driver analysis Physical counts and valuation Theoretical against actual usage Waste, transfers and returns Lot and stock conservation Demand-aware reorder proposals Usable-stock checks Supplier-substitution economics Inventory runway and stockouts Inventory exception queues Purchase drafts Supplier-master versions Order acknowledgments Partial receiving Cancellations and fees Payment and refund records Time and payroll evidence Payroll correction replay Payroll-close workflow Staff capacity and effort Internal staff handoffs Workforce continuity models Attendance and sales-change signals Workload and operating-state measures Neural survival-model comparison Evidence-led investigations Persistent work and dependencies Scheduled local execution Eleven operating playbooks Restaurant AI assistant Follow-up conversations Time-budgeted next steps Completion and execution records Crisis and recovery readiness Owner reports and decision queues Daily sales forecasts Forecast uncertainty Forward outcome monitoring Chronological backtesting Time-series foundation models Bounded model adaptation Temporal graph transformers Mechanistic restaurant models Model diagnostics and stress tests Budgeted model discovery Contextual decision learning Operator-preference learning Completion learning Economic-outcome learning Durable learning updates Policy evaluation Collective-learning transfer Temporal knowledge graph Business time and knowledge time Source lineage Permission-aware retrieval Source-bound data repair Freshness and conflict checks Cited calculations Replay and local recovery Campaign briefs and copy Editable creative drafts Creative review and export checks Review and reputation analysis Reservation and public context Advertising experiment workflow POS and labor connections Banking and cards through Plaid Payroll connections Supplier and invoice connections Connection health and replay Accounting, reservations, delivery and reviews Hosted subscriptions and external execution

No utility matches that search.

Every utility, by area

Pricing

$299 per location per month.

No contracts.

  1. The assistant. 100 evidence-grounded investigations a month, with follow-up conversations.
  2. Connected operating intelligence. Bookkeeping preparation, cost visibility, demand planning and cash scenarios where your source coverage supports them.
  3. Guided setup. Two setup sessions, standard supported imports and one agreed first workflow.
  4. Owner plus two managers, with configured permissions.
  5. A first-month value review. Completed work, your effort, and evidenced cost changes against the baseline.
Start a founding pilot Scale plans for groups opening by invitation.

No setup fee. Billing starts after the baseline is ready and you approve activation. Cancel before the next monthly renewal. Annual: $2,988 per location per year. The full package

Guardrails

Read-only by default. No money moves.

  • Read-only by default

    Every connection starts in read mode; dry run is the default for every adapter.

  • Approval before anything changes

    Posting journals, paying bills, placing orders and moving money always wait for a person.

  • A receipt for every decision

    Write-once, never rewritten, never deleted, binding the decision, the actor and the output hash.

  • A private workspace

    Content-addressed, encrypted evidence; employees pseudonymized; no guest record; raw rows never reach a model.

The security page

Connections

MarginVector OS connects to what you already run.

Point of sale

ToastSquareCloverSpotOnLightspeedTouchBistroRevelOracle MICROS SimphonyNCR AlohaAloha CloudPAR BrinkShift4SkyTab by Shift4Focus POSPOSitouchArryvedGoTabHungerRushLingaQuTabitSquirrelRezkuCake by Mad MobileEpos NowZonal

Banking and cards

Plaid: bank accounts, cards and transactions

Payroll and scheduling

GustoADPPaychexToast PayrollSquare Payroll7shifts

Suppliers and invoices

SyscoUS FoodsPerformance FoodserviceGordon Food ServiceBen E. KeithShamrock FoodsInvoice capture by emailInvoice capture by photo

Accounting

QuickBooks OnlineXeroSage IntacctNetSuiteRestaurant365

Reservations

OpenTableResySevenRooms

Delivery

DoorDashUber EatsGrubhub

Reviews

GoogleYelpTripadvisor

New connections are added on request. Every connection reads before it proposes, and MFA is always yours to complete.

Who built it

Built to run his own restaurant first.

MarginVector OS was built by a restaurant operator with a PhD in applied physics who manages AI model risk at a major bank. He owns and runs an independent full-service restaurant and built the platform to run it: nightly POS exports, weekly payroll, the books, inventory and advertising. The physics is in the restaurant model, dynamical systems applied to a dining room. The model-risk discipline is everywhere else: leakage gates, sealed holdouts, chronological backtests, promotion gates that fail closed, hash-chained ledgers and a critic that checks every answer against the evidence. No bank affiliation or endorsement is implied.

Questions

Before you ask.

Will it change anything without my approval?

No. Agents propose. Consequential actions wait for a person, and every decision leaves a receipt.

What does it connect to?

Your POS, your bank and card accounts through Plaid, payroll, supplier invoices, accounting, reservations, delivery and reviews. Every connection reads first, with an evidence hash on every record.

Does it replace my POS, my accountant or my payroll processor?

No. It is the intelligence and workflow layer around the records you already have. Draft journals go to your CPA for review. Payroll files go to your processor.

Is my data used to train shared models?

No. Your evidence stays in your private workspace, models that see your data are adapted for your restaurant alone, and your preferences are never pooled.

How long does setup take?

Four stages, each with a deliverable: a source check, connect and reconcile, configure and rehearse, then the first 30 days after activation with a value review.

What does the $299 include?

The assistant with 100 investigations a month, connected operating intelligence, guided setup with two sessions, owner plus two managers, and a first-month value review. Custom integrations, historical cleanup and provider pass-through fees are scoped separately.

Can I cancel?

Yes. No contracts. Cancel before the next monthly renewal, and billing starts only after you approve activation.

What about groups?

Scale plans for groups are opening by invitation. Write to us with your locations and your POS.

Start

Start a founding pilot.

$299 per location per month. No contracts. Billing starts after you approve activation.

See a 20-minute walkthrough