Scan, review, and let prices, margins, and inventory keep themselves current. Every AI output is reviewable; every automatic number is correctable.
Continuous capture built for speed: framing guides, a confirmation flash, and a live thumbnail of the last shot. The camera never blocks between photos — uploads and AI run in the background. An on-device quality gate (~5 ms) catches blurry or dark shots before they upload.
Four stages, each catching what the last can't: Classify (is this a purchase document, invoice or delivery note?), Extract (vendor, date, line items, per-line category and pack contents, with matching against your catalog), Validate (deterministic math cross-checks), and Review (a human approves in Triage).
The inbox where digitized documents wait. Review is side-by-side — the receipt photo next to editable line items, with amber warnings and flagged rows where the math didn't check out. Approve and prices roll; mark non-food bills as expenses; flip invoice ↔ delivery note if the AI misjudged. Automatic delivery-note ↔ invoice reconciliation catches billing errors.
The home screen answers "how is the restaurant doing?" at a glance: weekly expenses, revenue, food-cost % (against the 28–35% target band), and pantry value. Hand-rolled SVG charts: expenses vs revenue, the menu engineering matrix, price sparklines, and an ingredient Pareto. Every name links to its page.
Every dish costed live from your suppliers' rolling prices. The list shows target-vs-actual margin bars sorted worst-first — dishes needing attention float up. Each dish has a cost breakdown: each ingredient's share of the plate, the "why is this dish expensive" view.
A reference table, not a ledger: purchases (from scans) − sales (via recipes) = current stock, with zero daily input. Monthly, the chef walks the walk-in and trues up what they see. The grocery-list generator derives usage rate and suggests 7-day quantities, sorted by urgency and grouped by vendor — order by email or WhatsApp.
A directory that maintains itself — derived from approved invoices and tagged expenses, so it can never drift from reality. Each vendor shows spend, deliveries, last delivery, and derived tags: the categories they actually supply and the expenses you've paid them under. No manual tagging, ever.
A ✨ button opens Q&A over your own data: "what went up in price this week?". The context the model sees is filtered by the member's permissions, so a scan-only employee literally cannot ask about the money.
One restaurant is a shared workspace. The owner invites by email; on Google sign-in each member is attached with granular per-area permissions — scan (yes/no) and triage / menu / pantry / finance at none / read-only / edit. "Camera-only runner", "accountant with finance read-only", or "chef with everything but the money" are just permission combinations. Enforced server-side on every route. And it works the same with one location or many: on the Group plan the same permissions extend across your whole group with a consolidated view.
