Restaurant operating system
Restronik runs the whole restaurant — menu, orders, payments, tables, delivery and staff — for single venues and multi-branch brands. The part that usually takes days of typing is done by AI: it reads your printed menu, drafts your floor plan from photos, translates everything and shoots the food.
| HummusStarters | 6.50 | clear |
| Lentil SoupStarters | 5.00 | clear |
| Slow-Roasted LambMains | 24.00 | clear |
| Sea BassMains | null | check |
| Lemonade (1 L)Drinks | 9.50 | clear |
| Lemonade (0.5 L)Drinks | 5.50 | clear |
| Mint TeaDrinks | null | unreadable |
Illustration of the import preview. A market price and a smudged number come back empty and flagged — never guessed.
The problem
Before the first order can be rung up, someone has to retype the menu, draw every table, translate it for tourists and find a photo for each dish. Owners do it at midnight or pay an installer to. Restronik treats that setup work as the product's job.
AI features · built and working
Each one ends in a review screen, not a silent write. The owner confirms; the model never commits data on its own.
Point it at what the restaurant already has. It returns product, category, price, unit, size and description as rows the owner reviews before anything is saved.
Up to three eye-level photos of the same room become one top-down table plan, dropped onto the layout editor as unassigned tables.
Menu items, categories, modifiers, payment methods, service charges and surveys are filled column by column across the restaurant's menu languages.
Product and category images generated from the menu entry plus an optional direction from the owner, then applied to the catalogue from one studio screen.
How it's built
Restaurants run on thin margins and real money. Every AI call in Restronik is metered, bounded and reviewable before it touches a menu or a bill.
Credits are reserved before a model is called and settled against real token usage afterwards. If the balance can't cover the estimate, the call is never made.
Every call writes a usage event: feature, model, tokens, duration, outcome. One credit equals one micro-dollar of model cost, so pricing stays honest as models change.
Menus, web pages and photos are treated as data, never as instructions. Prompts are versioned and say so explicitly; imported pages can't steer the model.
Unreadable prices return empty and flagged. A room with too little evidence returns no plan. A wrong price on a menu costs more than a blank one.
Each capability sits behind its own contract with its own driver, limits and timeouts, so the best model for a job can be adopted without touching the product.
All AI work runs in background jobs with progress and retry, so a slow model never freezes the panel in the middle of service.
The platform around it
The AI features sit on a full multi-tenant platform, currently in development ahead of launch.
Organisation and restaurant panels, with brand-level defaults each branch can override.
Dine-in, takeaway, delivery with zone mapping, and self-service — each with its own settings.
Categories, products, modifiers, taxes and units, with multilingual content per restaurant.
Payment methods per channel, foreign-currency payment with live rates, customer deposits and open accounts.
Terminals, stations, printers and receipt templates, paired and licensed per branch.
Role-based permissions, staff invitations, a versioned public API and a five-language panel.
Roadmap · not built yet
Setup is where AI earns trust. The same metered, review-first foundation is designed to carry the features owners will use every day.
Plain-language questions over sales, stock and shifts — "which dishes lost margin this month?" — answered from the restaurant's own data.
A conversational helper that knows the restaurant's configuration and can walk an owner through a change instead of a settings hunt.
When importing from a web page, match each product to its existing photo so the catalogue arrives complete.